diff --git a/client/src/protoFleet/api/clients.ts b/client/src/protoFleet/api/clients.ts
index 6da8b36c1..638dd0145 100644
--- a/client/src/protoFleet/api/clients.ts
+++ b/client/src/protoFleet/api/clients.ts
@@ -11,6 +11,7 @@ import { ApiKeyService } from "@/protoFleet/api/generated/apikey/v1/apikey_pb";
 import { AuthService } from "@/protoFleet/api/generated/auth/v1/auth_pb";
 import { AuthzService } from "@/protoFleet/api/generated/authz/v1/authz_pb";
 import { BuildingService } from "@/protoFleet/api/generated/buildings/v1/buildings_pb";
+import { ChatService } from "@/protoFleet/api/generated/chat/v1/chat_pb";
 import { CurtailmentService } from "@/protoFleet/api/generated/curtailment/v1/curtailment_pb";
 import { DeviceSetService } from "@/protoFleet/api/generated/device_set/v1/device_set_pb";
 import { ErrorQueryService } from "@/protoFleet/api/generated/errors/v1/errors_pb";
@@ -51,6 +52,7 @@ const foremanImportClient = createClient(ForemanImportService, transport);
 const infrastructureClient = createClient(InfrastructureService, transport);
 const sitesClient = createClient(SiteService, transport);
 const buildingsClient = createClient(BuildingService, transport);
+const chatClient = createClient(ChatService, transport);
 const alertChannelClient = createClient(AlertChannelService, transport);
 const alertRuleClient = createClient(AlertRuleService, transport);
 const alertMaintenanceWindowClient = createClient(AlertMaintenanceWindowService, transport);
@@ -66,6 +68,7 @@ export {
   authClient,
   authzClient,
   buildingsClient,
+  chatClient,
   curtailmentClient,
   deviceSetClient,
   errorQueryClient,
diff --git a/client/src/protoFleet/api/generated/chat/v1/chat_pb.ts b/client/src/protoFleet/api/generated/chat/v1/chat_pb.ts
new file mode 100644
index 000000000..59fef2edb
--- /dev/null
+++ b/client/src/protoFleet/api/generated/chat/v1/chat_pb.ts
@@ -0,0 +1,734 @@
+// @generated by protoc-gen-es v2.13.0 with parameter "target=ts"
+// @generated from file chat/v1/chat.proto (package chat.v1, syntax proto3)
+/* eslint-disable */
+
+import type { GenEnum, GenFile, GenMessage, GenService } from "@bufbuild/protobuf/codegenv2";
+import { enumDesc, fileDesc, messageDesc, serviceDesc } from "@bufbuild/protobuf/codegenv2";
+import { file_buf_validate_validate } from "../../buf/validate/validate_pb";
+import type { Message } from "@bufbuild/protobuf";
+
+/**
+ * Describes the file chat/v1/chat.proto.
+ */
+export const file_chat_v1_chat: GenFile =
+  /*@__PURE__*/
+  fileDesc(
+    "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",
+    [file_buf_validate_validate],
+  );
+
+/**
+ * @generated from message chat.v1.LLMConfig
+ */
+export type LLMConfig = Message<"chat.v1.LLMConfig"> & {
+  /**
+   * @generated from field: chat.v1.AgentHarness harness = 1;
+   */
+  harness: AgentHarness;
+
+  /**
+   * @generated from field: chat.v1.LLMProvider provider = 2;
+   */
+  provider: LLMProvider;
+
+  /**
+   * @generated from field: bool has_api_key = 3;
+   */
+  hasApiKey: boolean;
+
+  /**
+   * @generated from field: string base_url = 4;
+   */
+  baseUrl: string;
+
+  /**
+   * @generated from field: string model = 5;
+   */
+  model: string;
+
+  /**
+   * @generated from field: string goose_base_url = 7;
+   */
+  gooseBaseUrl: string;
+
+  /**
+   * @generated from field: bool has_goose_secret = 8;
+   */
+  hasGooseSecret: boolean;
+
+  /**
+   * @generated from field: bool configured = 9;
+   */
+  configured: boolean;
+};
+
+/**
+ * Describes the message chat.v1.LLMConfig.
+ * Use `create(LLMConfigSchema)` to create a new message.
+ */
+export const LLMConfigSchema: GenMessage<LLMConfig> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 0);
+
+/**
+ * @generated from message chat.v1.GetLLMConfigRequest
+ */
+export type GetLLMConfigRequest = Message<"chat.v1.GetLLMConfigRequest"> & {};
+
+/**
+ * Describes the message chat.v1.GetLLMConfigRequest.
+ * Use `create(GetLLMConfigRequestSchema)` to create a new message.
+ */
+export const GetLLMConfigRequestSchema: GenMessage<GetLLMConfigRequest> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 1);
+
+/**
+ * @generated from message chat.v1.GetLLMConfigResponse
+ */
+export type GetLLMConfigResponse = Message<"chat.v1.GetLLMConfigResponse"> & {
+  /**
+   * @generated from field: chat.v1.LLMConfig config = 1;
+   */
+  config?: LLMConfig | undefined;
+};
+
+/**
+ * Describes the message chat.v1.GetLLMConfigResponse.
+ * Use `create(GetLLMConfigResponseSchema)` to create a new message.
+ */
+export const GetLLMConfigResponseSchema: GenMessage<GetLLMConfigResponse> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 2);
+
+/**
+ * @generated from message chat.v1.DiscoverModelsRequest
+ */
+export type DiscoverModelsRequest = Message<"chat.v1.DiscoverModelsRequest"> & {
+  /**
+   * @generated from field: chat.v1.LLMProvider provider = 1;
+   */
+  provider: LLMProvider;
+
+  /**
+   * Write-only credential used for this discovery request. It is not saved.
+   *
+   * @generated from field: string api_key = 2;
+   */
+  apiKey: string;
+
+  /**
+   * @generated from field: string base_url = 3;
+   */
+  baseUrl: string;
+
+  /**
+   * Uses the encrypted credential already saved for this provider. This cannot
+   * be combined with api_key.
+   *
+   * @generated from field: bool use_stored_api_key = 4;
+   */
+  useStoredApiKey: boolean;
+};
+
+/**
+ * Describes the message chat.v1.DiscoverModelsRequest.
+ * Use `create(DiscoverModelsRequestSchema)` to create a new message.
+ */
+export const DiscoverModelsRequestSchema: GenMessage<DiscoverModelsRequest> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 3);
+
+/**
+ * @generated from message chat.v1.AvailableModel
+ */
+export type AvailableModel = Message<"chat.v1.AvailableModel"> & {
+  /**
+   * @generated from field: string id = 1;
+   */
+  id: string;
+
+  /**
+   * @generated from field: string display_name = 2;
+   */
+  displayName: string;
+};
+
+/**
+ * Describes the message chat.v1.AvailableModel.
+ * Use `create(AvailableModelSchema)` to create a new message.
+ */
+export const AvailableModelSchema: GenMessage<AvailableModel> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 4);
+
+/**
+ * @generated from message chat.v1.DiscoverModelsResponse
+ */
+export type DiscoverModelsResponse = Message<"chat.v1.DiscoverModelsResponse"> & {
+  /**
+   * @generated from field: repeated chat.v1.AvailableModel models = 1;
+   */
+  models: AvailableModel[];
+};
+
+/**
+ * Describes the message chat.v1.DiscoverModelsResponse.
+ * Use `create(DiscoverModelsResponseSchema)` to create a new message.
+ */
+export const DiscoverModelsResponseSchema: GenMessage<DiscoverModelsResponse> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 5);
+
+/**
+ * @generated from message chat.v1.UpdateLLMConfigRequest
+ */
+export type UpdateLLMConfigRequest = Message<"chat.v1.UpdateLLMConfigRequest"> & {
+  /**
+   * @generated from field: chat.v1.AgentHarness harness = 1;
+   */
+  harness: AgentHarness;
+
+  /**
+   * @generated from field: chat.v1.LLMProvider provider = 2;
+   */
+  provider: LLMProvider;
+
+  /**
+   * Write-only. An empty value preserves the existing key unless
+   * clear_api_key is true.
+   *
+   * @generated from field: string api_key = 3;
+   */
+  apiKey: string;
+
+  /**
+   * @generated from field: string base_url = 4;
+   */
+  baseUrl: string;
+
+  /**
+   * @generated from field: string model = 5;
+   */
+  model: string;
+
+  /**
+   * @generated from field: string goose_base_url = 7;
+   */
+  gooseBaseUrl: string;
+
+  /**
+   * Write-only shared secret for a remote Goose ACP endpoint.
+   *
+   * @generated from field: string goose_secret = 8;
+   */
+  gooseSecret: string;
+
+  /**
+   * @generated from field: bool clear_api_key = 9;
+   */
+  clearApiKey: boolean;
+
+  /**
+   * @generated from field: bool clear_goose_secret = 10;
+   */
+  clearGooseSecret: boolean;
+};
+
+/**
+ * Describes the message chat.v1.UpdateLLMConfigRequest.
+ * Use `create(UpdateLLMConfigRequestSchema)` to create a new message.
+ */
+export const UpdateLLMConfigRequestSchema: GenMessage<UpdateLLMConfigRequest> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 6);
+
+/**
+ * @generated from message chat.v1.UpdateLLMConfigResponse
+ */
+export type UpdateLLMConfigResponse = Message<"chat.v1.UpdateLLMConfigResponse"> & {
+  /**
+   * @generated from field: chat.v1.LLMConfig config = 1;
+   */
+  config?: LLMConfig | undefined;
+};
+
+/**
+ * Describes the message chat.v1.UpdateLLMConfigResponse.
+ * Use `create(UpdateLLMConfigResponseSchema)` to create a new message.
+ */
+export const UpdateLLMConfigResponseSchema: GenMessage<UpdateLLMConfigResponse> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 7);
+
+/**
+ * @generated from message chat.v1.ChatTurn
+ */
+export type ChatTurn = Message<"chat.v1.ChatTurn"> & {
+  /**
+   * @generated from field: chat.v1.ChatRole role = 1;
+   */
+  role: ChatRole;
+
+  /**
+   * @generated from field: string content = 2;
+   */
+  content: string;
+};
+
+/**
+ * Describes the message chat.v1.ChatTurn.
+ * Use `create(ChatTurnSchema)` to create a new message.
+ */
+export const ChatTurnSchema: GenMessage<ChatTurn> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 8);
+
+/**
+ * @generated from message chat.v1.SendMessageRequest
+ */
+export type SendMessageRequest = Message<"chat.v1.SendMessageRequest"> & {
+  /**
+   * @generated from field: string conversation_id = 1;
+   */
+  conversationId: string;
+
+  /**
+   * @generated from field: string content = 2;
+   */
+  content: string;
+
+  /**
+   * Conversations remain client-side. Supplying prior turns keeps the server
+   * stateless while preserving multi-turn context.
+   *
+   * @generated from field: repeated chat.v1.ChatTurn history = 3;
+   */
+  history: ChatTurn[];
+};
+
+/**
+ * Describes the message chat.v1.SendMessageRequest.
+ * Use `create(SendMessageRequestSchema)` to create a new message.
+ */
+export const SendMessageRequestSchema: GenMessage<SendMessageRequest> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 9);
+
+/**
+ * @generated from message chat.v1.TextDelta
+ */
+export type TextDelta = Message<"chat.v1.TextDelta"> & {
+  /**
+   * @generated from field: string content = 1;
+   */
+  content: string;
+};
+
+/**
+ * Describes the message chat.v1.TextDelta.
+ * Use `create(TextDeltaSchema)` to create a new message.
+ */
+export const TextDeltaSchema: GenMessage<TextDelta> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 10);
+
+/**
+ * @generated from message chat.v1.ToolCall
+ */
+export type ToolCall = Message<"chat.v1.ToolCall"> & {
+  /**
+   * @generated from field: string id = 1;
+   */
+  id: string;
+
+  /**
+   * @generated from field: string name = 2;
+   */
+  name: string;
+
+  /**
+   * @generated from field: string summary = 3;
+   */
+  summary: string;
+};
+
+/**
+ * Describes the message chat.v1.ToolCall.
+ * Use `create(ToolCallSchema)` to create a new message.
+ */
+export const ToolCallSchema: GenMessage<ToolCall> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 11);
+
+/**
+ * @generated from message chat.v1.ToolResult
+ */
+export type ToolResult = Message<"chat.v1.ToolResult"> & {
+  /**
+   * @generated from field: string id = 1;
+   */
+  id: string;
+
+  /**
+   * @generated from field: string name = 2;
+   */
+  name: string;
+
+  /**
+   * @generated from field: bool success = 3;
+   */
+  success: boolean;
+
+  /**
+   * @generated from field: string summary = 4;
+   */
+  summary: string;
+
+  /**
+   * @generated from field: bool cancelled = 5;
+   */
+  cancelled: boolean;
+};
+
+/**
+ * Describes the message chat.v1.ToolResult.
+ * Use `create(ToolResultSchema)` to create a new message.
+ */
+export const ToolResultSchema: GenMessage<ToolResult> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 12);
+
+/**
+ * @generated from message chat.v1.ToolConfirmationDetail
+ */
+export type ToolConfirmationDetail = Message<"chat.v1.ToolConfirmationDetail"> & {
+  /**
+   * @generated from field: string label = 1;
+   */
+  label: string;
+
+  /**
+   * @generated from field: string value = 2;
+   */
+  value: string;
+};
+
+/**
+ * Describes the message chat.v1.ToolConfirmationDetail.
+ * Use `create(ToolConfirmationDetailSchema)` to create a new message.
+ */
+export const ToolConfirmationDetailSchema: GenMessage<ToolConfirmationDetail> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 13);
+
+/**
+ * @generated from message chat.v1.ToolConfirmationRequired
+ */
+export type ToolConfirmationRequired = Message<"chat.v1.ToolConfirmationRequired"> & {
+  /**
+   * Opaque, short-lived identifier bound to the current operator and org.
+   *
+   * @generated from field: string confirmation_id = 1;
+   */
+  confirmationId: string;
+
+  /**
+   * @generated from field: string tool_call_id = 2;
+   */
+  toolCallId: string;
+
+  /**
+   * @generated from field: string tool_name = 3;
+   */
+  toolName: string;
+
+  /**
+   * @generated from field: string title = 4;
+   */
+  title: string;
+
+  /**
+   * @generated from field: string description = 5;
+   */
+  description: string;
+
+  /**
+   * @generated from field: string confirm_label = 6;
+   */
+  confirmLabel: string;
+
+  /**
+   * @generated from field: repeated chat.v1.ToolConfirmationDetail details = 7;
+   */
+  details: ToolConfirmationDetail[];
+};
+
+/**
+ * Describes the message chat.v1.ToolConfirmationRequired.
+ * Use `create(ToolConfirmationRequiredSchema)` to create a new message.
+ */
+export const ToolConfirmationRequiredSchema: GenMessage<ToolConfirmationRequired> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 14);
+
+/**
+ * @generated from message chat.v1.ResolveToolConfirmationRequest
+ */
+export type ResolveToolConfirmationRequest = Message<"chat.v1.ResolveToolConfirmationRequest"> & {
+  /**
+   * @generated from field: string confirmation_id = 1;
+   */
+  confirmationId: string;
+
+  /**
+   * @generated from field: chat.v1.ToolConfirmationDecision decision = 2;
+   */
+  decision: ToolConfirmationDecision;
+};
+
+/**
+ * Describes the message chat.v1.ResolveToolConfirmationRequest.
+ * Use `create(ResolveToolConfirmationRequestSchema)` to create a new message.
+ */
+export const ResolveToolConfirmationRequestSchema: GenMessage<ResolveToolConfirmationRequest> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 15);
+
+/**
+ * @generated from message chat.v1.ResolveToolConfirmationResponse
+ */
+export type ResolveToolConfirmationResponse = Message<"chat.v1.ResolveToolConfirmationResponse"> & {};
+
+/**
+ * Describes the message chat.v1.ResolveToolConfirmationResponse.
+ * Use `create(ResolveToolConfirmationResponseSchema)` to create a new message.
+ */
+export const ResolveToolConfirmationResponseSchema: GenMessage<ResolveToolConfirmationResponse> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 16);
+
+/**
+ * @generated from message chat.v1.Done
+ */
+export type Done = Message<"chat.v1.Done"> & {
+  /**
+   * @generated from field: string finish_reason = 1;
+   */
+  finishReason: string;
+};
+
+/**
+ * Describes the message chat.v1.Done.
+ * Use `create(DoneSchema)` to create a new message.
+ */
+export const DoneSchema: GenMessage<Done> = /*@__PURE__*/ messageDesc(file_chat_v1_chat, 17);
+
+/**
+ * @generated from message chat.v1.SendMessageResponse
+ */
+export type SendMessageResponse = Message<"chat.v1.SendMessageResponse"> & {
+  /**
+   * @generated from field: string message_id = 1;
+   */
+  messageId: string;
+
+  /**
+   * @generated from oneof chat.v1.SendMessageResponse.event
+   */
+  event:
+    | {
+        /**
+         * @generated from field: chat.v1.TextDelta text_delta = 2;
+         */
+        value: TextDelta;
+        case: "textDelta";
+      }
+    | {
+        /**
+         * @generated from field: chat.v1.ToolCall tool_call = 3;
+         */
+        value: ToolCall;
+        case: "toolCall";
+      }
+    | {
+        /**
+         * @generated from field: chat.v1.ToolResult tool_result = 4;
+         */
+        value: ToolResult;
+        case: "toolResult";
+      }
+    | {
+        /**
+         * @generated from field: chat.v1.Done done = 5;
+         */
+        value: Done;
+        case: "done";
+      }
+    | {
+        /**
+         * @generated from field: chat.v1.ToolConfirmationRequired confirmation_required = 6;
+         */
+        value: ToolConfirmationRequired;
+        case: "confirmationRequired";
+      }
+    | { case: undefined; value?: undefined };
+};
+
+/**
+ * Describes the message chat.v1.SendMessageResponse.
+ * Use `create(SendMessageResponseSchema)` to create a new message.
+ */
+export const SendMessageResponseSchema: GenMessage<SendMessageResponse> =
+  /*@__PURE__*/
+  messageDesc(file_chat_v1_chat, 18);
+
+/**
+ * @generated from enum chat.v1.AgentHarness
+ */
+export enum AgentHarness {
+  /**
+   * @generated from enum value: AGENT_HARNESS_UNSPECIFIED = 0;
+   */
+  UNSPECIFIED = 0,
+
+  /**
+   * @generated from enum value: AGENT_HARNESS_NATIVE = 1;
+   */
+  NATIVE = 1,
+
+  /**
+   * @generated from enum value: AGENT_HARNESS_GOOSE = 2;
+   */
+  GOOSE = 2,
+}
+
+/**
+ * Describes the enum chat.v1.AgentHarness.
+ */
+export const AgentHarnessSchema: GenEnum<AgentHarness> = /*@__PURE__*/ enumDesc(file_chat_v1_chat, 0);
+
+/**
+ * @generated from enum chat.v1.LLMProvider
+ */
+export enum LLMProvider {
+  /**
+   * @generated from enum value: LLM_PROVIDER_UNSPECIFIED = 0;
+   */
+  LLM_PROVIDER_UNSPECIFIED = 0,
+
+  /**
+   * @generated from enum value: LLM_PROVIDER_OPENAI = 1;
+   */
+  LLM_PROVIDER_OPENAI = 1,
+
+  /**
+   * @generated from enum value: LLM_PROVIDER_ANTHROPIC = 2;
+   */
+  LLM_PROVIDER_ANTHROPIC = 2,
+
+  /**
+   * @generated from enum value: LLM_PROVIDER_OLLAMA = 3;
+   */
+  LLM_PROVIDER_OLLAMA = 3,
+
+  /**
+   * @generated from enum value: LLM_PROVIDER_CUSTOM = 4;
+   */
+  LLM_PROVIDER_CUSTOM = 4,
+}
+
+/**
+ * Describes the enum chat.v1.LLMProvider.
+ */
+export const LLMProviderSchema: GenEnum<LLMProvider> = /*@__PURE__*/ enumDesc(file_chat_v1_chat, 1);
+
+/**
+ * @generated from enum chat.v1.ChatRole
+ */
+export enum ChatRole {
+  /**
+   * @generated from enum value: CHAT_ROLE_UNSPECIFIED = 0;
+   */
+  UNSPECIFIED = 0,
+
+  /**
+   * @generated from enum value: CHAT_ROLE_USER = 1;
+   */
+  USER = 1,
+
+  /**
+   * @generated from enum value: CHAT_ROLE_ASSISTANT = 2;
+   */
+  ASSISTANT = 2,
+}
+
+/**
+ * Describes the enum chat.v1.ChatRole.
+ */
+export const ChatRoleSchema: GenEnum<ChatRole> = /*@__PURE__*/ enumDesc(file_chat_v1_chat, 2);
+
+/**
+ * @generated from enum chat.v1.ToolConfirmationDecision
+ */
+export enum ToolConfirmationDecision {
+  /**
+   * @generated from enum value: TOOL_CONFIRMATION_DECISION_UNSPECIFIED = 0;
+   */
+  UNSPECIFIED = 0,
+
+  /**
+   * @generated from enum value: TOOL_CONFIRMATION_DECISION_APPROVE = 1;
+   */
+  APPROVE = 1,
+
+  /**
+   * @generated from enum value: TOOL_CONFIRMATION_DECISION_CANCEL = 2;
+   */
+  CANCEL = 2,
+}
+
+/**
+ * Describes the enum chat.v1.ToolConfirmationDecision.
+ */
+export const ToolConfirmationDecisionSchema: GenEnum<ToolConfirmationDecision> =
+  /*@__PURE__*/
+  enumDesc(file_chat_v1_chat, 3);
+
+/**
+ * ChatService owns the organization-scoped BYOLLM configuration and the
+ * streaming agent conversation boundary used by the Proto Fleet web client.
+ *
+ * @generated from service chat.v1.ChatService
+ */
+export const ChatService: GenService<{
+  /**
+   * @generated from rpc chat.v1.ChatService.GetLLMConfig
+   */
+  getLLMConfig: {
+    methodKind: "unary";
+    input: typeof GetLLMConfigRequestSchema;
+    output: typeof GetLLMConfigResponseSchema;
+  };
+  /**
+   * @generated from rpc chat.v1.ChatService.DiscoverModels
+   */
+  discoverModels: {
+    methodKind: "unary";
+    input: typeof DiscoverModelsRequestSchema;
+    output: typeof DiscoverModelsResponseSchema;
+  };
+  /**
+   * @generated from rpc chat.v1.ChatService.UpdateLLMConfig
+   */
+  updateLLMConfig: {
+    methodKind: "unary";
+    input: typeof UpdateLLMConfigRequestSchema;
+    output: typeof UpdateLLMConfigResponseSchema;
+  };
+  /**
+   * @generated from rpc chat.v1.ChatService.SendMessage
+   */
+  sendMessage: {
+    methodKind: "server_streaming";
+    input: typeof SendMessageRequestSchema;
+    output: typeof SendMessageResponseSchema;
+  };
+  /**
+   * ResolveToolConfirmation approves or cancels a pending write requested by
+   * SendMessage. The original stream remains open and resumes after this RPC.
+   *
+   * @generated from rpc chat.v1.ChatService.ResolveToolConfirmation
+   */
+  resolveToolConfirmation: {
+    methodKind: "unary";
+    input: typeof ResolveToolConfirmationRequestSchema;
+    output: typeof ResolveToolConfirmationResponseSchema;
+  };
+}> = /*@__PURE__*/ serviceDesc(file_chat_v1_chat, 0);
diff --git a/client/src/protoFleet/components/AppLayout/AppLayout.test.tsx b/client/src/protoFleet/components/AppLayout/AppLayout.test.tsx
index 3c093bbf3..8b7343885 100644
--- a/client/src/protoFleet/components/AppLayout/AppLayout.test.tsx
+++ b/client/src/protoFleet/components/AppLayout/AppLayout.test.tsx
@@ -34,6 +34,11 @@ vi.mock("@/protoFleet/components/PageHeader", () => ({
   default: () => <div>Page header</div>,
 }));
 
+vi.mock("@/protoFleet/features/aiChat", () => ({
+  ChatFab: () => <div data-testid="ai-chat-fab" />,
+  ChatPanel: () => <div data-testid="ai-chat-panel" />,
+}));
+
 vi.mock("@/protoFleet/components/PageHeader/useSchedulePillData", () => ({
   useSchedulePillData: () => mockUseSchedulePillData(),
 }));
@@ -184,6 +189,19 @@ describe("AppLayout", () => {
     expect(screen.getByText("Body content").parentElement).toHaveClass("top-[calc(theme(spacing.1)*12)]");
   });
 
+  it("hides the floating Minerbot launcher on the dedicated Minerbot page", () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <AppLayout>
+          <div>Body content</div>
+        </AppLayout>
+      </MemoryRouter>,
+    );
+
+    expect(screen.queryByTestId("ai-chat-fab")).not.toBeInTheDocument();
+    expect(screen.queryByTestId("ai-chat-panel")).not.toBeInTheDocument();
+  });
+
   it("uses the two-widget phone content offset when all three header widgets are visible", () => {
     mockUseReactiveLocalStorage.mockReturnValue([true, vi.fn()]);
     mockUseCurtailmentPillData.mockReturnValue({ activeEvent: activeCurtailmentEvent });
@@ -251,4 +269,32 @@ describe("AppLayout", () => {
 
     expect(screen.getByText("Body content").parentElement).toHaveClass("phone:top-[calc(theme(spacing.1)*12)]");
   });
+
+  it("shows Minerbot chat when the operator can read the fleet", () => {
+    render(
+      <MemoryRouter>
+        <AppLayout>
+          <div>Body content</div>
+        </AppLayout>
+      </MemoryRouter>,
+    );
+
+    expect(screen.getByTestId("ai-chat-fab")).toBeInTheDocument();
+    expect(screen.getByTestId("ai-chat-panel")).toBeInTheDocument();
+  });
+
+  it("hides Minerbot chat when the operator cannot read the fleet", () => {
+    vi.mocked(useHasPermission).mockImplementation((permission) => permission !== "fleet:read");
+
+    render(
+      <MemoryRouter>
+        <AppLayout>
+          <div>Body content</div>
+        </AppLayout>
+      </MemoryRouter>,
+    );
+
+    expect(screen.queryByTestId("ai-chat-fab")).not.toBeInTheDocument();
+    expect(screen.queryByTestId("ai-chat-panel")).not.toBeInTheDocument();
+  });
 });
diff --git a/client/src/protoFleet/components/AppLayout/AppLayout.tsx b/client/src/protoFleet/components/AppLayout/AppLayout.tsx
index 516b4533a..5c0e6e217 100644
--- a/client/src/protoFleet/components/AppLayout/AppLayout.tsx
+++ b/client/src/protoFleet/components/AppLayout/AppLayout.tsx
@@ -1,4 +1,5 @@
 import { ReactNode, useEffect, useRef, useState } from "react";
+import { useLocation } from "react-router-dom";
 import clsx from "clsx";
 
 import NavigationMenu from "../NavigationMenu";
@@ -16,6 +17,7 @@ import {
 import { useCurtailmentPillData } from "@/protoFleet/components/PageHeader/useCurtailmentPillData";
 import { useSchedulePillData } from "@/protoFleet/components/PageHeader/useSchedulePillData";
 import { primaryNavItems } from "@/protoFleet/config/navItems";
+import { ChatFab, ChatPanel } from "@/protoFleet/features/aiChat";
 import { usePageBackground } from "@/protoFleet/hooks/usePageBackground";
 import { useHasPermission } from "@/protoFleet/store";
 import { Menu } from "@/shared/assets/icons";
@@ -29,6 +31,7 @@ type Props = {
 
 const AppLayoutContent = ({ children, hideShellHeader = false }: Props) => {
   const [isMenuOpen, setIsMenuOpen] = useState(false);
+  const { pathname } = useLocation();
   const { bgClass } = usePageBackground();
   const { isPhone, isTablet } = useWindowDimensions();
   const [dismissedSetup] = useReactiveLocalStorage<boolean>("completeSetupDismissed");
@@ -36,6 +39,7 @@ const AppLayoutContent = ({ children, hideShellHeader = false }: Props) => {
   const { activeEvent: activeCurtailmentEvent } = useCurtailmentPillData();
   const hasDismissedSetup = Boolean(dismissedSetup);
   const canReadCurtailment = useHasPermission("curtailment:read");
+  const canReadFleet = useHasPermission("fleet:read");
   const hasVisibleCurtailmentPill = activeCurtailmentEvent !== null && canReadCurtailment;
   const headerWidgetCount = getVisibleHeaderWidgetCount({
     hasDismissedSetup,
@@ -46,6 +50,7 @@ const AppLayoutContent = ({ children, hideShellHeader = false }: Props) => {
   const phoneRowWidgetCount = getPhoneHeaderWidgetRowCount(headerWidgetCount, inlineFirstPhoneWidget);
   const stackPhoneWidgets = shouldStackPhoneHeaderWidgets(headerWidgetCount);
 
+  const isMinerbotPage = pathname === "/minerbot";
   const showDetailMenuTrigger = hideShellHeader && (isPhone || isTablet) && !isMenuOpen;
   const showPhoneWidgets = !hideShellHeader && isPhone && phoneRowWidgetCount > 0;
 
@@ -115,6 +120,14 @@ const AppLayoutContent = ({ children, hideShellHeader = false }: Props) => {
       >
         {children}
       </div>
+
+      {canReadFleet && !isMinerbotPage ? (
+        <>
+          {/* AI Chat: FAB launcher + slide-up panel */}
+          <ChatFab />
+          <ChatPanel />
+        </>
+      ) : null}
     </div>
   );
 };
diff --git a/client/src/protoFleet/config/navItems.test.ts b/client/src/protoFleet/config/navItems.test.ts
index 2ccf3f4da..1e622e2f3 100644
--- a/client/src/protoFleet/config/navItems.test.ts
+++ b/client/src/protoFleet/config/navItems.test.ts
@@ -1,7 +1,7 @@
 import { describe, expect, it } from "vitest";
 
 import { getSettingsLandingPath, isNavItemAllowedByPermissions, primaryNavItems, secondaryNavItems } from "./navItems";
-import { LightningAlt } from "@/shared/assets/icons";
+import { AIStroked, LightningAlt } from "@/shared/assets/icons";
 
 describe("primaryNavItems", () => {
   it("shows Energy above Activity with the electric icon", () => {
@@ -51,6 +51,18 @@ describe("primaryNavItems", () => {
     expect(home?.requiredPermission).toBeUndefined();
     expect(home?.requiredAnyPermission).toBeUndefined();
   });
+
+  it("adds Minerbot as a fleet-readable primary destination before Settings", () => {
+    const labels = primaryNavItems.map((item) => item.label);
+    const minerbot = primaryNavItems.find((item) => item.label === "Minerbot");
+
+    expect(minerbot).toMatchObject({
+      path: "/minerbot",
+      icon: AIStroked,
+      requiredPermission: "fleet:read",
+    });
+    expect(labels.indexOf("Minerbot")).toBe(labels.indexOf("Settings") - 1);
+  });
 });
 
 describe("secondaryNavItems", () => {
@@ -100,6 +112,18 @@ describe("secondaryNavItems", () => {
     );
   });
 
+  it("gives agent configuration its own Admin destination", () => {
+    expect(secondaryNavItems).toContainEqual(
+      expect.objectContaining({
+        path: "/settings/agents",
+        label: "Agents",
+        parent: "/settings",
+        section: "Admin",
+        requiredPermission: "apikey:manage",
+      }),
+    );
+  });
+
   it("folds role management into the Team destination", () => {
     expect(secondaryNavItems).toContainEqual(
       expect.objectContaining({
diff --git a/client/src/protoFleet/config/navItems.ts b/client/src/protoFleet/config/navItems.ts
index fb2ad915e..d9b471302 100644
--- a/client/src/protoFleet/config/navItems.ts
+++ b/client/src/protoFleet/config/navItems.ts
@@ -1,6 +1,6 @@
 import { type ReactNode } from "react";
 
-import { Activity, Fleet, Groups, Home, IconProps, LightningAlt, Settings } from "@/shared/assets/icons";
+import { Activity, AIStroked, Fleet, Groups, Home, IconProps, LightningAlt, Settings } from "@/shared/assets/icons";
 
 // Runtime-gated features: an entry tagged with one is shown only when the server
 // reports the feature enabled (see SecondaryNavigation). Distinct from
@@ -102,6 +102,12 @@ export const primaryNavItems: NavItem[] = [
     requiredPermission: "activity:read",
     scopable: true,
   },
+  {
+    path: "/minerbot",
+    label: "Minerbot",
+    icon: AIStroked,
+    requiredPermission: "fleet:read",
+  },
   {
     path: "/settings",
     label: "Settings",
@@ -191,6 +197,13 @@ export const secondaryNavItems: SecondaryNavItem[] = [
     // surface is usable so role-only admins are not stranded after the merge.
     requiredAnyPermission: ["user:read", "role:manage"],
   },
+  {
+    path: "/settings/agents",
+    label: "Agents",
+    parent: "/settings",
+    section: "Admin",
+    requiredPermission: "apikey:manage",
+  },
   {
     path: "/settings/integrations",
     label: "Integrations",
diff --git a/client/src/protoFleet/features/aiChat/AgentActivityStatus.test.tsx b/client/src/protoFleet/features/aiChat/AgentActivityStatus.test.tsx
new file mode 100644
index 000000000..441c89db5
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/AgentActivityStatus.test.tsx
@@ -0,0 +1,31 @@
+import { render, screen } from "@testing-library/react";
+import { describe, expect, test } from "vitest";
+
+import AgentActivityStatus from "./AgentActivityStatus";
+import type { AgentActivity } from "./types";
+
+const activity = (status: AgentActivity["status"], summary: string): AgentActivity => ({
+  id: "activity-1",
+  status,
+  summary,
+  timestamp: new Date("2026-07-16T12:00:00Z"),
+  sequence: 1,
+});
+
+describe("AgentActivityStatus", () => {
+  test("renders compact running, completed, and unsuccessful states", () => {
+    const { rerender } = render(<AgentActivityStatus activity={activity("running", "Checking fleet health")} />);
+
+    expect(screen.getByRole("img", { name: "In progress" })).toBeInTheDocument();
+    expect(screen.getByText("Checking fleet health")).toBeInTheDocument();
+    expect(screen.getByTestId("agent-activity-status").tagName).toBe("DIV");
+
+    rerender(<AgentActivityStatus activity={activity("completed", "Read state for 14 miners")} />);
+    expect(screen.getByRole("img", { name: "Completed" })).toBeInTheDocument();
+    expect(screen.getByText("Read state for 14 miners")).toBeInTheDocument();
+
+    rerender(<AgentActivityStatus activity={activity("failed", "Unable to read this fleet data")} />);
+    expect(screen.getByRole("img", { name: "Couldn't complete" })).toBeInTheDocument();
+    expect(screen.getByText("Unable to read this fleet data")).toBeInTheDocument();
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/AgentActivityStatus.tsx b/client/src/protoFleet/features/aiChat/AgentActivityStatus.tsx
new file mode 100644
index 000000000..c369ad8c9
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/AgentActivityStatus.tsx
@@ -0,0 +1,42 @@
+import clsx from "clsx";
+
+import type { AgentActivity } from "./types";
+import { Checkmark, DismissTiny } from "@/shared/assets/icons";
+import ProgressCircular from "@/shared/components/ProgressCircular";
+
+const STATUS_LABELS: Record<AgentActivity["status"], string> = {
+  running: "In progress",
+  completed: "Completed",
+  failed: "Couldn't complete",
+  cancelled: "Cancelled",
+};
+
+interface AgentActivityStatusProps {
+  activity: AgentActivity;
+}
+
+const AgentActivityStatus = ({ activity }: AgentActivityStatusProps) => (
+  <div
+    className="flex min-w-0 items-center gap-2 py-1 text-200 text-text-primary-50"
+    data-testid="agent-activity-status"
+  >
+    <span
+      aria-label={STATUS_LABELS[activity.status]}
+      className={clsx("flex size-4 shrink-0 items-center justify-center", {
+        "text-text-success": activity.status === "completed",
+        "text-intent-info-fill": activity.status === "running",
+        "text-text-critical": activity.status === "failed",
+        "text-text-primary-30": activity.status === "cancelled",
+      })}
+      role="img"
+    >
+      {activity.status === "completed" ? <Checkmark width="w-3.5" /> : null}
+      {activity.status === "running" ? <ProgressCircular indeterminate size={14} /> : null}
+      {activity.status === "failed" ? <DismissTiny width="w-3" /> : null}
+      {activity.status === "cancelled" ? <DismissTiny width="w-3" /> : null}
+    </span>
+    <span className="min-w-0 break-words">{activity.summary}</span>
+  </div>
+);
+
+export default AgentActivityStatus;
diff --git a/client/src/protoFleet/features/aiChat/ChatConversation.tsx b/client/src/protoFleet/features/aiChat/ChatConversation.tsx
new file mode 100644
index 000000000..16b5a2c96
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatConversation.tsx
@@ -0,0 +1,123 @@
+import { type ReactNode, type Ref } from "react";
+import clsx from "clsx";
+
+import AgentActivityStatus from "./AgentActivityStatus";
+import ChatMessageContent from "./ChatMessageContent";
+import ToolConfirmationCard from "./ToolConfirmationCard";
+import type { ToolConfirmation } from "./types";
+import type { MinerbotConversationItem } from "./useMinerbotConversation";
+import { Alert } from "@/shared/assets/icons";
+import Callout from "@/shared/components/Callout";
+
+type ChatConversationProps = {
+  conversationEndRef?: Ref<HTMLDivElement>;
+  conversationItems: MinerbotConversationItem[];
+  header?: ReactNode;
+  hasPendingConfirmation: boolean;
+  isStreaming: boolean;
+  layout?: "panel" | "fullScreen";
+  onOpenSettings: () => void;
+  onResolveConfirmation: (confirmation: ToolConfirmation, decision: "approve" | "cancel") => void;
+  streamError: string;
+  streamingContent: string;
+};
+
+const ChatConversation = ({
+  conversationEndRef,
+  conversationItems,
+  header,
+  hasPendingConfirmation,
+  isStreaming,
+  layout = "panel",
+  onOpenSettings,
+  onResolveConfirmation,
+  streamError,
+  streamingContent,
+}: ChatConversationProps) => {
+  const isFullScreen = layout === "fullScreen";
+
+  return (
+    <div
+      aria-label="Conversation"
+      aria-live="polite"
+      className={clsx("flex flex-col", isFullScreen ? "mx-auto w-full max-w-[800px] gap-6" : "gap-4")}
+    >
+      {header ? <div className="border-b border-border-5 pb-5">{header}</div> : null}
+      {conversationItems.map((item) =>
+        item.kind === "activity" ? (
+          <AgentActivityStatus key={`activity-${item.activity.id}`} activity={item.activity} />
+        ) : item.kind === "confirmation" ? (
+          <ToolConfirmationCard
+            key={`confirmation-${item.confirmation.id}`}
+            confirmation={item.confirmation}
+            onResolve={onResolveConfirmation}
+          />
+        ) : (
+          <div
+            key={`message-${item.message.id}`}
+            className={`flex ${item.message.role === "user" ? "justify-end" : "justify-start"}`}
+          >
+            <div
+              className={clsx(
+                "text-300",
+                item.message.role === "user"
+                  ? "max-w-[85%] rounded-2xl rounded-br-md bg-core-primary-fill px-4 py-3 whitespace-pre-wrap text-text-contrast"
+                  : isFullScreen
+                    ? "w-full text-text-primary"
+                    : "max-w-full rounded-2xl rounded-bl-md bg-core-primary-5 px-4 py-3 text-text-primary",
+              )}
+            >
+              {item.message.role === "assistant" ? (
+                <ChatMessageContent content={item.message.content} />
+              ) : (
+                item.message.content
+              )}
+            </div>
+          </div>
+        ),
+      )}
+      {isStreaming && streamingContent ? (
+        <div className="flex justify-start">
+          <div
+            className={clsx(
+              "text-300 text-text-primary",
+              isFullScreen ? "w-full" : "max-w-full rounded-2xl rounded-bl-md bg-core-primary-5 px-4 py-3",
+            )}
+          >
+            <ChatMessageContent content={streamingContent} />
+          </div>
+        </div>
+      ) : null}
+      {isStreaming && !streamingContent && !hasPendingConfirmation ? (
+        <div className="flex justify-start" aria-label="Minerbot is responding" role="status">
+          <div
+            className={clsx(
+              "flex items-center gap-1",
+              isFullScreen ? "py-2" : "rounded-2xl rounded-bl-md bg-core-primary-5 px-4 py-4",
+            )}
+          >
+            {[0, 1, 2].map((dot) => (
+              <span
+                key={dot}
+                className="size-1.5 animate-pulse rounded-full bg-text-primary-50"
+                style={{ animationDelay: `${dot * 120}ms` }}
+              />
+            ))}
+          </div>
+        </div>
+      ) : null}
+      {streamError ? (
+        <Callout
+          buttonOnClick={onOpenSettings}
+          buttonText="Open Minerbot settings"
+          intent="danger"
+          prefixIcon={<Alert />}
+          title={streamError}
+        />
+      ) : null}
+      <div ref={conversationEndRef} />
+    </div>
+  );
+};
+
+export default ChatConversation;
diff --git a/client/src/protoFleet/features/aiChat/ChatFab.tsx b/client/src/protoFleet/features/aiChat/ChatFab.tsx
new file mode 100644
index 000000000..2a1cd5fab
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatFab.tsx
@@ -0,0 +1,34 @@
+import { AnimatePresence, motion } from "motion/react";
+
+import { useChatStore } from "./useChatStore";
+import { AI } from "@/shared/assets/icons";
+
+const ChatFab = () => {
+  const isOpen = useChatStore((state) => state.isOpen);
+  const open = useChatStore((state) => state.open);
+
+  return (
+    <AnimatePresence>
+      {!isOpen ? (
+        <motion.button
+          type="button"
+          aria-controls="ai-chat-panel"
+          aria-expanded={false}
+          aria-label="Open Minerbot"
+          className="fixed right-6 bottom-6 z-[45] flex size-14 items-center justify-center rounded-full bg-core-accent-fill text-text-base-contrast-static shadow-300 outline-none hover:opacity-80 focus-visible:ring-2 focus-visible:ring-core-primary-fill focus-visible:ring-offset-2 focus-visible:ring-offset-surface-base phone:right-4 phone:bottom-4"
+          data-testid="ai-chat-fab"
+          initial={{ opacity: 0, scale: 0.8 }}
+          animate={{ opacity: 1, scale: 1 }}
+          exit={{ opacity: 0, scale: 0.8 }}
+          transition={{ duration: 0.2 }}
+          whileTap={{ scale: 0.94 }}
+          onClick={open}
+        >
+          <AI innerShadow={false} width="w-6" />
+        </motion.button>
+      ) : null}
+    </AnimatePresence>
+  );
+};
+
+export default ChatFab;
diff --git a/client/src/protoFleet/features/aiChat/ChatInput.tsx b/client/src/protoFleet/features/aiChat/ChatInput.tsx
new file mode 100644
index 000000000..00200508a
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatInput.tsx
@@ -0,0 +1,64 @@
+import { FormEvent, KeyboardEvent, useState } from "react";
+
+import { ArrowUp } from "@/shared/assets/icons";
+import Button, { variants } from "@/shared/components/Button";
+
+type ChatInputProps = {
+  disabled?: boolean;
+  onSend: (content: string) => void;
+};
+
+const ChatInput = ({ disabled = false, onSend }: ChatInputProps) => {
+  const [content, setContent] = useState("");
+  const canSend = content.trim().length > 0 && !disabled;
+
+  const submit = () => {
+    const nextMessage = content.trim();
+    if (!nextMessage || disabled) return;
+    onSend(nextMessage);
+    setContent("");
+  };
+
+  const handleSubmit = (event: FormEvent<HTMLFormElement>) => {
+    event.preventDefault();
+    submit();
+  };
+
+  const handleKeyDown = (event: KeyboardEvent<HTMLTextAreaElement>) => {
+    if (event.key === "Enter" && !event.shiftKey) {
+      event.preventDefault();
+      submit();
+    }
+  };
+
+  return (
+    <form className="rounded-2xl border border-border-10 bg-surface-base p-2 shadow-100" onSubmit={handleSubmit}>
+      <label className="sr-only" htmlFor="ai-chat-input">
+        Message Minerbot
+      </label>
+      <div className="flex items-end gap-2">
+        <textarea
+          id="ai-chat-input"
+          aria-label="Message Minerbot"
+          className="max-h-24 min-h-10 min-w-0 flex-1 resize-none bg-transparent px-2 py-2 text-300 text-text-primary outline-none placeholder:text-text-primary-50 pointer-coarse:text-400"
+          disabled={disabled}
+          onChange={(event) => setContent(event.target.value)}
+          onKeyDown={handleKeyDown}
+          placeholder="Ask about your fleet…"
+          rows={1}
+          value={content}
+        />
+        <Button
+          ariaLabel="Send message"
+          className="!h-10 !w-10 shrink-0 !rounded-full !p-0"
+          disabled={!canSend}
+          onClick={submit}
+          prefixIcon={<ArrowUp width="w-5" />}
+          variant={variants.primary}
+        />
+      </div>
+    </form>
+  );
+};
+
+export default ChatInput;
diff --git a/client/src/protoFleet/features/aiChat/ChatMessageContent.test.tsx b/client/src/protoFleet/features/aiChat/ChatMessageContent.test.tsx
new file mode 100644
index 000000000..70a65ed6c
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatMessageContent.test.tsx
@@ -0,0 +1,42 @@
+import { render, screen, within } from "@testing-library/react";
+import { describe, expect, test } from "vitest";
+
+import ChatMessageContent from "./ChatMessageContent";
+
+describe("ChatMessageContent", () => {
+  test("renders Markdown data tables as structured HTML", () => {
+    render(
+      <ChatMessageContent
+        content={`Fleet health
+
+| Miner state | Count |
+| --- | ---: |
+| Total | 14 |
+| Hashing | 0 |
+| Offline | 14 |
+
+All miners are offline.`}
+      />,
+    );
+
+    const table = screen.getByRole("table");
+    expect(table).toHaveClass("text-300");
+    expect(table).not.toHaveClass("text-200");
+    expect(within(table).getByRole("columnheader", { name: "Miner state" })).toBeInTheDocument();
+    expect(within(table).getByRole("columnheader", { name: "Miner state" })).toHaveClass("text-emphasis-300");
+    expect(within(table).getByRole("columnheader", { name: "Miner state" })).not.toHaveClass("text-emphasis-200");
+    expect(within(table).getByRole("columnheader", { name: "Count" })).toHaveClass("text-right");
+    const offlineRow = within(table).getByRole("cell", { name: "Offline" }).closest("tr");
+    expect(offlineRow).not.toBeNull();
+    expect(within(offlineRow as HTMLTableRowElement).getByRole("cell", { name: "14" })).toHaveClass("tabular-nums");
+    expect(screen.getByText("All miners are offline.")).toBeInTheDocument();
+    expect(screen.queryByText(/---:/)).not.toBeInTheDocument();
+  });
+
+  test("keeps ordinary answers as whitespace-preserving text", () => {
+    const { container } = render(<ChatMessageContent content={"Offline: 14\nNo sites configured."} />);
+
+    expect(screen.queryByRole("table")).not.toBeInTheDocument();
+    expect(container.querySelector(".whitespace-pre-wrap")).toHaveTextContent("Offline: 14 No sites configured.");
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/ChatMessageContent.tsx b/client/src/protoFleet/features/aiChat/ChatMessageContent.tsx
new file mode 100644
index 000000000..4679c8d8c
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatMessageContent.tsx
@@ -0,0 +1,177 @@
+import { Fragment, type ReactNode } from "react";
+
+type Alignment = "left" | "center" | "right";
+
+type TextBlock = {
+  kind: "text";
+  content: string;
+};
+
+type TableBlock = {
+  kind: "table";
+  headers: string[];
+  rows: string[][];
+  alignments: Alignment[];
+};
+
+type ContentBlock = TextBlock | TableBlock;
+
+const TABLE_DIVIDER = /^:?-{3,}:?$/;
+const NUMERIC_VALUE = /^[-+]?(?:\d{1,3}(?:,\d{3})*|\d+)(?:\.\d+)?(?:\s*(?:%|[A-Za-z]+(?:\/[A-Za-z]+)?))?$/;
+
+const parseTableRow = (line: string): string[] | null => {
+  const trimmed = line.trim();
+  if (!trimmed.includes("|")) return null;
+
+  const row = trimmed.replace(/^\|/, "").replace(/\|$/, "");
+  const cells: string[] = [];
+  let cell = "";
+  let escaped = false;
+
+  for (const character of row) {
+    if (escaped) {
+      cell += character === "|" ? "|" : `\\${character}`;
+      escaped = false;
+    } else if (character === "\\") {
+      escaped = true;
+    } else if (character === "|") {
+      cells.push(cell.trim());
+      cell = "";
+    } else {
+      cell += character;
+    }
+  }
+  if (escaped) cell += "\\";
+  cells.push(cell.trim());
+
+  return cells.length >= 2 ? cells : null;
+};
+
+const dividerAlignment = (divider: string): Alignment => {
+  if (divider.startsWith(":") && divider.endsWith(":")) return "center";
+  if (divider.endsWith(":")) return "right";
+  return "left";
+};
+
+const parseContent = (content: string): ContentBlock[] => {
+  const lines = content.replace(/\r\n/g, "\n").split("\n");
+  const blocks: ContentBlock[] = [];
+  let textLines: string[] = [];
+
+  const flushText = () => {
+    const text = textLines.join("\n").trim();
+    if (text) blocks.push({ kind: "text", content: text });
+    textLines = [];
+  };
+
+  for (let lineIndex = 0; lineIndex < lines.length;) {
+    const headers = parseTableRow(lines[lineIndex]);
+    const dividers = lineIndex + 1 < lines.length ? parseTableRow(lines[lineIndex + 1]) : null;
+    const isTable =
+      headers !== null &&
+      dividers !== null &&
+      headers.length === dividers.length &&
+      dividers.every((divider) => TABLE_DIVIDER.test(divider));
+
+    if (!isTable || !headers || !dividers) {
+      textLines.push(lines[lineIndex]);
+      lineIndex += 1;
+      continue;
+    }
+
+    flushText();
+    const rows: string[][] = [];
+    lineIndex += 2;
+    while (lineIndex < lines.length) {
+      const row = parseTableRow(lines[lineIndex]);
+      if (!row || row.length !== headers.length) break;
+      rows.push(row);
+      lineIndex += 1;
+    }
+
+    const alignments = dividers.map((divider, columnIndex) => {
+      const explicitAlignment = dividerAlignment(divider);
+      if (explicitAlignment !== "left") return explicitAlignment;
+      const numericColumn = rows.length > 0 && rows.every((row) => NUMERIC_VALUE.test(row[columnIndex]));
+      return numericColumn ? "right" : "left";
+    });
+    blocks.push({ kind: "table", headers, rows, alignments });
+  }
+
+  flushText();
+  return blocks;
+};
+
+const renderInlineText = (content: string): ReactNode =>
+  content.split(/(\*\*[^*\n]+\*\*)/g).map((part, index) =>
+    part.startsWith("**") && part.endsWith("**") ? (
+      <strong key={`${part}-${index}`} className="font-semibold">
+        {part.slice(2, -2)}
+      </strong>
+    ) : (
+      <Fragment key={`${part}-${index}`}>{part}</Fragment>
+    ),
+  );
+
+const alignmentClass = (alignment: Alignment) => {
+  if (alignment === "right") return "text-right tabular-nums";
+  if (alignment === "center") return "text-center";
+  return "text-left";
+};
+
+interface ChatMessageContentProps {
+  content: string;
+}
+
+const ChatMessageContent = ({ content }: ChatMessageContentProps) => {
+  const blocks = parseContent(content);
+
+  return (
+    <div className="flex min-w-0 flex-col gap-3">
+      {blocks.map((block, blockIndex) =>
+        block.kind === "text" ? (
+          <div key={`text-${blockIndex}`} className="break-words whitespace-pre-wrap">
+            {renderInlineText(block.content)}
+          </div>
+        ) : (
+          <div
+            key={`table-${blockIndex}`}
+            className="max-w-full overflow-x-auto rounded-xl border border-border-5 bg-surface-base"
+          >
+            <table className="w-full min-w-max border-collapse text-300">
+              <thead className="bg-core-primary-5">
+                <tr>
+                  {block.headers.map((header, columnIndex) => (
+                    <th
+                      key={`${header}-${columnIndex}`}
+                      scope="col"
+                      className={`border-b border-border-5 px-3 py-2 text-emphasis-300 whitespace-nowrap ${alignmentClass(block.alignments[columnIndex])}`}
+                    >
+                      {renderInlineText(header)}
+                    </th>
+                  ))}
+                </tr>
+              </thead>
+              <tbody>
+                {block.rows.map((row, rowIndex) => (
+                  <tr key={`${row.join("-")}-${rowIndex}`} className="border-t border-border-5 first:border-t-0">
+                    {row.map((cell, columnIndex) => (
+                      <td
+                        key={`${cell}-${columnIndex}`}
+                        className={`px-3 py-2 align-top ${alignmentClass(block.alignments[columnIndex])}`}
+                      >
+                        {renderInlineText(cell)}
+                      </td>
+                    ))}
+                  </tr>
+                ))}
+              </tbody>
+            </table>
+          </div>
+        ),
+      )}
+    </div>
+  );
+};
+
+export default ChatMessageContent;
diff --git a/client/src/protoFleet/features/aiChat/ChatPanel.test.tsx b/client/src/protoFleet/features/aiChat/ChatPanel.test.tsx
new file mode 100644
index 000000000..3e1031fa9
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatPanel.test.tsx
@@ -0,0 +1,199 @@
+import { MemoryRouter } from "react-router-dom";
+import { act, fireEvent, render, screen } from "@testing-library/react";
+import { afterEach, beforeEach, describe, expect, test, vi } from "vitest";
+
+import ChatPanel from "./ChatPanel";
+import { useChatStore } from "./useChatStore";
+
+const mocks = vi.hoisted(() => ({
+  sendMessage: vi.fn(),
+  resolveToolConfirmation: vi.fn(),
+}));
+const scrollIntoViewMock = vi.fn();
+
+vi.mock("@/protoFleet/api/clients", () => ({
+  chatClient: mocks,
+}));
+
+describe("ChatPanel", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    scrollIntoViewMock.mockClear();
+    Element.prototype.scrollIntoView = scrollIntoViewMock;
+    act(() => {
+      useChatStore.getState().clearMessages();
+      useChatStore.getState().open();
+    });
+  });
+
+  afterEach(() => {
+    act(() => {
+      useChatStore.getState().close();
+      useChatStore.getState().clearMessages();
+    });
+  });
+
+  test("shows the live agent empty state without demo activity", () => {
+    render(
+      <MemoryRouter>
+        <ChatPanel />
+      </MemoryRouter>,
+    );
+
+    expect(screen.getByText("Beta")).toBeInTheDocument();
+    expect(screen.getByRole("heading", { name: "What would you like to know?" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "Summarize fleet health" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "Find profit risks today" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "What needs attention first?" })).toBeInTheDocument();
+    expect(screen.queryByText("Demo data")).not.toBeInTheDocument();
+    expect(screen.queryByText("RebootBot")).not.toBeInTheDocument();
+  });
+
+  test("renders assistant data tables as structured content", () => {
+    act(() => {
+      useChatStore.getState().addMessage("assistant", "| Miner state | Count |\n| --- | ---: |\n| Offline | 14 |");
+    });
+
+    render(
+      <MemoryRouter>
+        <ChatPanel />
+      </MemoryRouter>,
+    );
+
+    expect(screen.getByRole("table")).toBeInTheDocument();
+    expect(screen.getByRole("columnheader", { name: "Miner state" })).toBeInTheDocument();
+    expect(screen.getByRole("cell", { name: "Offline" })).toBeInTheDocument();
+  });
+
+  test("places compact tool statuses between the request and response", () => {
+    act(() => {
+      useChatStore.getState().addMessage("user", "Summarize fleet health");
+      useChatStore.getState().beginToolActivity("call-1", "Checking fleet health");
+      useChatStore.getState().finishToolActivity("call-1", true, "Read state for 14 miners");
+      useChatStore.getState().addMessage("assistant", "All 14 miners are offline.");
+    });
+
+    render(
+      <MemoryRouter>
+        <ChatPanel />
+      </MemoryRouter>,
+    );
+
+    const transcript = screen.getByLabelText("Conversation").textContent ?? "";
+    const requestIndex = transcript.indexOf("Summarize fleet health");
+    const statusIndex = transcript.indexOf("Read state for 14 miners");
+    const responseIndex = transcript.indexOf("All 14 miners are offline.");
+
+    expect(requestIndex).toBeGreaterThanOrEqual(0);
+    expect(statusIndex).toBeGreaterThan(requestIndex);
+    expect(responseIndex).toBeGreaterThan(statusIndex);
+    expect(screen.getByTestId("agent-activity-status")).not.toHaveClass("rounded-2xl");
+    expect(screen.getByTestId("ai-chat-panel-scroll-area")).toHaveClass("min-h-0", "overflow-y-auto", "scroll-pb-4");
+    expect(scrollIntoViewMock).toHaveBeenLastCalledWith({ behavior: "smooth", block: "end", inline: "nearest" });
+  });
+
+  test("cancels and ignores an in-flight response when starting a new chat", async () => {
+    let releaseStream = () => {};
+    let finishStream = () => {};
+    const streamReleased = new Promise<void>((resolve) => {
+      releaseStream = resolve;
+    });
+    const streamFinished = new Promise<void>((resolve) => {
+      finishStream = resolve;
+    });
+    let requestSignal: AbortSignal | undefined;
+    mocks.sendMessage.mockImplementation((_request, options) => {
+      requestSignal = options?.signal;
+      return (async function* () {
+        try {
+          await streamReleased;
+          yield { event: { case: "textDelta", value: { content: "Stale response" } } };
+        } finally {
+          finishStream();
+        }
+      })();
+    });
+
+    render(
+      <MemoryRouter>
+        <ChatPanel />
+      </MemoryRouter>,
+    );
+
+    fireEvent.click(screen.getByRole("button", { name: "Summarize fleet health" }));
+    fireEvent.click(await screen.findByRole("button", { name: "New chat" }));
+
+    expect(requestSignal?.aborted).toBe(true);
+    await act(async () => {
+      releaseStream();
+      await streamFinished;
+    });
+
+    expect(screen.queryByText("Stale response")).not.toBeInTheDocument();
+    expect(useChatStore.getState().messages).toEqual([]);
+    expect(useChatStore.getState().agentActivities).toEqual([]);
+    expect(useChatStore.getState().isStreaming).toBe(false);
+  });
+
+  test("keeps a write paused until the operator confirms it", async () => {
+    let releaseConfirmation = () => {};
+    const confirmationResolved = new Promise<void>((resolve) => {
+      releaseConfirmation = resolve;
+    });
+    mocks.resolveToolConfirmation.mockImplementation(async () => {
+      releaseConfirmation();
+      return {};
+    });
+    mocks.sendMessage.mockImplementation(() =>
+      (async function* () {
+        yield { event: { case: "toolCall", value: { id: "call-1", summary: "Preparing site creation" } } };
+        yield {
+          event: {
+            case: "confirmationRequired",
+            value: {
+              confirmationId: "confirmation-1",
+              toolCallId: "call-1",
+              title: "Create this site?",
+              description: "Minerbot will add a new site to your fleet.",
+              confirmLabel: "Create site",
+              details: [{ label: "Name", value: "North" }],
+            },
+          },
+        };
+        await confirmationResolved;
+        yield {
+          event: {
+            case: "toolResult",
+            value: { id: "call-1", success: true, cancelled: false, summary: 'Created site "North"' },
+          },
+        };
+        yield { event: { case: "textDelta", value: { content: "North was created." } } };
+      })(),
+    );
+
+    render(
+      <MemoryRouter>
+        <ChatPanel />
+      </MemoryRouter>,
+    );
+
+    fireEvent.change(screen.getByRole("textbox", { name: "Message Minerbot" }), {
+      target: { value: "Create a site named North" },
+    });
+    fireEvent.click(screen.getByRole("button", { name: "Send message" }));
+
+    expect(await screen.findByRole("heading", { name: "Create this site?" })).toBeInTheDocument();
+    expect(screen.getByText("North")).toBeInTheDocument();
+    expect(screen.queryByText("North was created.")).not.toBeInTheDocument();
+
+    fireEvent.click(screen.getByRole("button", { name: "Create site" }));
+
+    expect(mocks.resolveToolConfirmation).toHaveBeenCalledWith({
+      confirmationId: "confirmation-1",
+      decision: 1,
+    });
+    expect(await screen.findByText("North was created.")).toBeInTheDocument();
+    expect(screen.getByText("Approved")).toBeInTheDocument();
+    expect(screen.getByText('Created site "North"')).toBeInTheDocument();
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/ChatPanel.tsx b/client/src/protoFleet/features/aiChat/ChatPanel.tsx
new file mode 100644
index 000000000..6b675b907
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatPanel.tsx
@@ -0,0 +1,144 @@
+import { AnimatePresence, motion } from "motion/react";
+import { useCallback, useEffect, useRef } from "react";
+
+import ChatConversation from "./ChatConversation";
+import ChatInput from "./ChatInput";
+import { useChatStore } from "./useChatStore";
+import useMinerbotConversation from "./useMinerbotConversation";
+import { Dismiss } from "@/shared/assets/icons";
+import Button, { sizes as buttonSizes, variants as buttonVariants } from "@/shared/components/Button";
+import Chip from "@/shared/components/Chip";
+import Header from "@/shared/components/Header";
+
+const ChatPanel = () => {
+  const isOpen = useChatStore((state) => state.isOpen);
+  const close = useChatStore((state) => state.close);
+  const {
+    chatContext,
+    conversationItems,
+    hasConversation,
+    hasPendingConfirmation,
+    isStreaming,
+    openSettings,
+    resolveConfirmation,
+    sendMessage,
+    startNewChat,
+    streamError,
+    streamingContent,
+  } = useMinerbotConversation();
+  const conversationEndRef = useRef<HTMLDivElement>(null);
+
+  const handleOpenSettings = useCallback(() => {
+    close();
+    openSettings();
+  }, [close, openSettings]);
+
+  useEffect(() => {
+    if (!isOpen) return;
+    conversationEndRef.current?.scrollIntoView({ behavior: "smooth", block: "end", inline: "nearest" });
+  }, [conversationItems, isOpen, isStreaming, streamError, streamingContent]);
+
+  useEffect(() => {
+    if (!isOpen) return;
+    const handleEscape = (event: globalThis.KeyboardEvent) => {
+      if (event.key === "Escape") close();
+    };
+    window.addEventListener("keydown", handleEscape);
+    return () => window.removeEventListener("keydown", handleEscape);
+  }, [close, isOpen]);
+
+  return (
+    <AnimatePresence>
+      {isOpen ? (
+        <motion.section
+          id="ai-chat-panel"
+          aria-label="Minerbot chat"
+          className="fixed right-6 bottom-6 z-[45] flex max-h-[min(760px,calc(100dvh-48px))] w-[min(calc(100vw-theme(spacing.12)),480px)] flex-col overflow-hidden rounded-3xl border border-border-5 bg-surface-elevated-base text-text-primary shadow-300 phone:right-0 phone:bottom-0 phone:left-0 phone:max-h-[calc(100dvh-12px)] phone:w-auto phone:rounded-b-none"
+          data-testid="ai-chat-panel"
+          initial={{ opacity: 0, y: 32, scale: 0.98 }}
+          animate={{ opacity: 1, y: 0, scale: 1 }}
+          exit={{ opacity: 0, y: 32, scale: 0.98 }}
+          transition={{ duration: 0.25, ease: [0.47, 0, 0.23, 1] }}
+        >
+          <header className="shrink-0 border-b border-border-5 px-5 py-4">
+            <Header
+              title="Minerbot"
+              titleSize="text-heading-200"
+              icon={<Dismiss />}
+              iconAriaLabel="Close Minerbot"
+              iconOnClick={close}
+              inline
+              centerButton
+              stackButtonsOnPhone={false}
+              buttonSize={buttonSizes.compact}
+              buttons={
+                hasConversation
+                  ? [
+                      {
+                        text: "New chat",
+                        variant: buttonVariants.secondary,
+                        onClick: startNewChat,
+                      },
+                    ]
+                  : undefined
+              }
+              buttonsWrapperClassName="shrink-0"
+            >
+              {hasConversation ? null : (
+                <div className="shrink-0">
+                  <Chip>Beta</Chip>
+                </div>
+              )}
+            </Header>
+          </header>
+
+          <div
+            className="min-h-0 flex-1 scroll-pb-4 overflow-y-auto overscroll-contain px-5 pt-5 pb-4"
+            data-testid="ai-chat-panel-scroll-area"
+          >
+            {hasConversation ? (
+              <ChatConversation
+                conversationEndRef={conversationEndRef}
+                conversationItems={conversationItems}
+                hasPendingConfirmation={hasPendingConfirmation}
+                isStreaming={isStreaming}
+                onOpenSettings={handleOpenSettings}
+                onResolveConfirmation={resolveConfirmation}
+                streamError={streamError}
+                streamingContent={streamingContent}
+              />
+            ) : (
+              <div className="flex flex-col gap-6">
+                <div>
+                  <h2 className="text-heading-200 text-text-primary">What would you like to know?</h2>
+                  <p className="mt-1 text-300 text-text-primary-50">{chatContext.description}</p>
+                </div>
+
+                <div aria-label="Suggested prompts" className="flex flex-wrap gap-2">
+                  {chatContext.suggestions.map((suggestion) => (
+                    <Button
+                      key={suggestion.label}
+                      size={buttonSizes.compact}
+                      text={suggestion.label}
+                      onClick={() => sendMessage(suggestion.label)}
+                      variant={buttonVariants.secondary}
+                    />
+                  ))}
+                </div>
+              </div>
+            )}
+          </div>
+
+          <div className="shrink-0 border-t border-border-5 bg-surface-elevated-base px-5 pt-4 pb-5 phone:pb-[max(20px,env(safe-area-inset-bottom))]">
+            <ChatInput disabled={isStreaming} onSend={sendMessage} />
+            <p className="mt-2 text-center text-200 text-text-primary-30">
+              Minerbot asks for confirmation before making changes.
+            </p>
+          </div>
+        </motion.section>
+      ) : null}
+    </AnimatePresence>
+  );
+};
+
+export default ChatPanel;
diff --git a/client/src/protoFleet/features/aiChat/ChatShell.stories.tsx b/client/src/protoFleet/features/aiChat/ChatShell.stories.tsx
new file mode 100644
index 000000000..f5346462c
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ChatShell.stories.tsx
@@ -0,0 +1,166 @@
+import { useEffect } from "react";
+import type { Meta, StoryObj } from "@storybook/react";
+
+import ChatFab from "./ChatFab";
+import ChatPanel from "./ChatPanel";
+import { useChatStore } from "./useChatStore";
+
+type Scenario = "closed" | "empty" | "operationalReview" | "streaming";
+
+interface ChatShellStoryProps {
+  scenario: Scenario;
+}
+
+const seedScenario = (scenario: Scenario) => {
+  const store = useChatStore.getState();
+  store.clearMessages();
+  store.resetStream();
+  store.setStreaming(false);
+  store.open();
+
+  if (scenario === "closed") {
+    store.close();
+    return;
+  }
+
+  if (scenario === "empty") return;
+
+  store.addMessage("user", "Which firmware updates should I prioritize today?");
+  store.beginToolActivity("scan-firmware", "Checking firmware versions across the fleet");
+  store.finishToolActivity("scan-firmware", true, "Found 42 miners behind the target firmware");
+  store.addMessage(
+    "assistant",
+    "42 miners are behind the target firmware. Prioritize Building 2, Rack B, where stale firmware overlaps with a 6.8% hashrate drop.\n\n| Batch | Miners | Expected lift |\n| --- | ---: | ---: |\n| Building 2 / Rack B | 18 | 6.8% |\n| Building 1 / Rack D | 14 | 3.1% |\n| Yard units | 10 | 1.4% |",
+  );
+
+  if (scenario === "streaming") {
+    store.beginToolActivity("compare-market", "Comparing power price, pool fees, and hashprice");
+    store.setStreaming(true);
+    store.appendStreamingContent("I would hold curtailment for 2 hours and keep high-efficiency racks online first.");
+    return;
+  }
+
+  store.addToolConfirmation({
+    id: "firmware-update-confirmation",
+    toolCallId: "scan-firmware",
+    title: "Review firmware update?",
+    description: "Minerbot will prepare a staged update for the affected miners before any change is applied.",
+    confirmLabel: "Review update",
+    details: [
+      { label: "Scope", value: "42 miners" },
+      { label: "Batch size", value: "10 miners" },
+      { label: "Maintenance window", value: "Today, 11:00 PM" },
+    ],
+  });
+};
+
+const ChatShellStory = ({ scenario }: ChatShellStoryProps) => {
+  useEffect(() => {
+    seedScenario(scenario);
+
+    return () => {
+      const store = useChatStore.getState();
+      store.close();
+      store.clearMessages();
+      store.resetStream();
+      store.setStreaming(false);
+    };
+  }, [scenario]);
+
+  return (
+    <div className="min-h-screen bg-surface-10 p-8 text-text-primary">
+      <div className="mx-auto grid max-w-[1120px] gap-5">
+        <header className="flex items-center justify-between">
+          <div>
+            <h1 className="text-heading-300">Fleet dashboard</h1>
+            <p className="mt-1 text-300 text-text-primary-50">Cedar Creek site</p>
+          </div>
+          <div className="rounded-full bg-intent-success-10 px-3 py-1 text-emphasis-200 text-text-success">
+            94% online
+          </div>
+        </header>
+
+        <div className="grid gap-4 tablet:grid-cols-3">
+          {[
+            ["Hashrate", "812 PH/s", "+2.8%"],
+            ["Power draw", "26.4 MW", "-1.1%"],
+            ["Revenue pace", "$43.2K", "+4.5%"],
+          ].map(([label, value, trend]) => (
+            <section key={label} className="rounded-lg border border-border-5 bg-surface-base p-4">
+              <p className="text-200 text-text-primary-50">{label}</p>
+              <div className="mt-2 flex items-end justify-between">
+                <p className="text-heading-300">{value}</p>
+                <p className="text-emphasis-200 text-text-success">{trend}</p>
+              </div>
+            </section>
+          ))}
+        </div>
+
+        <section className="rounded-lg border border-border-5 bg-surface-base p-5">
+          <div className="flex items-center justify-between">
+            <h2 className="text-heading-200">Rack health</h2>
+            <p className="text-200 text-text-primary-50">Live operating view</p>
+          </div>
+          <div className="mt-5 grid grid-cols-12 gap-2">
+            {Array.from({ length: 72 }, (_, index) => {
+              const stateClass =
+                index % 29 === 0
+                  ? "bg-intent-critical-fill"
+                  : index % 11 === 0
+                    ? "bg-intent-warning-fill"
+                    : "bg-intent-success-fill";
+              return <div key={index} className={`h-8 rounded-sm ${stateClass}`} />;
+            })}
+          </div>
+        </section>
+      </div>
+
+      <ChatFab />
+      <ChatPanel />
+    </div>
+  );
+};
+
+const meta = {
+  title: "Proto Fleet/AI Chat/Chat Shell",
+  component: ChatShellStory,
+  args: {
+    scenario: "operationalReview",
+  },
+  argTypes: {
+    scenario: {
+      control: "select",
+      options: ["closed", "empty", "operationalReview", "streaming"],
+    },
+  },
+  parameters: {
+    layout: "fullscreen",
+  },
+} satisfies Meta<typeof ChatShellStory>;
+
+export default meta;
+type Story = StoryObj<typeof meta>;
+
+export const OperationalReview: Story = {
+  args: {
+    scenario: "operationalReview",
+  },
+};
+
+export const Empty: Story = {
+  args: {
+    scenario: "empty",
+  },
+};
+
+export const ClosedFab: Story = {
+  args: {
+    scenario: "closed",
+  },
+};
+
+export const Streaming: Story = {
+  args: {
+    scenario: "streaming",
+  },
+};
diff --git a/client/src/protoFleet/features/aiChat/ToolConfirmationCard.test.tsx b/client/src/protoFleet/features/aiChat/ToolConfirmationCard.test.tsx
new file mode 100644
index 000000000..fc8029abb
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ToolConfirmationCard.test.tsx
@@ -0,0 +1,47 @@
+import { fireEvent, render, screen } from "@testing-library/react";
+import { describe, expect, test, vi } from "vitest";
+
+import ToolConfirmationCard from "./ToolConfirmationCard";
+import type { ToolConfirmation } from "./types";
+
+const pendingConfirmation: ToolConfirmation = {
+  id: "confirmation-1",
+  toolCallId: "call-1",
+  title: "Move 2 miners?",
+  description: "Minerbot will move these miners.",
+  confirmLabel: "Move miners",
+  details: [
+    { label: "Destination rack ID", value: "21" },
+    { label: "Miners", value: "miner-a, miner-b" },
+  ],
+  status: "pending",
+  sequence: 2,
+};
+
+describe("ToolConfirmationCard", () => {
+  test("shows exact action details and exposes approve and cancel decisions", () => {
+    const onResolve = vi.fn();
+    render(<ToolConfirmationCard confirmation={pendingConfirmation} onResolve={onResolve} />);
+
+    expect(screen.getByRole("heading", { name: "Move 2 miners?" })).toBeInTheDocument();
+    expect(screen.getByText("miner-a, miner-b")).toBeInTheDocument();
+
+    fireEvent.click(screen.getByRole("button", { name: "Cancel" }));
+    expect(onResolve).toHaveBeenCalledWith(pendingConfirmation, "cancel");
+
+    fireEvent.click(screen.getByRole("button", { name: "Move miners" }));
+    expect(onResolve).toHaveBeenCalledWith(pendingConfirmation, "approve");
+  });
+
+  test("locks both decisions while the response is being submitted", () => {
+    render(
+      <ToolConfirmationCard
+        confirmation={{ ...pendingConfirmation, status: "submitting", decision: "approve" }}
+        onResolve={vi.fn()}
+      />,
+    );
+
+    expect(screen.getByRole("button", { name: "Cancel" })).toBeDisabled();
+    expect(screen.getByRole("button", { name: "Submitting" })).toBeDisabled();
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/ToolConfirmationCard.tsx b/client/src/protoFleet/features/aiChat/ToolConfirmationCard.tsx
new file mode 100644
index 000000000..411c308d9
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/ToolConfirmationCard.tsx
@@ -0,0 +1,87 @@
+import type { ToolConfirmation } from "./types";
+import { Alert } from "@/shared/assets/icons";
+import { variants } from "@/shared/components/Button";
+import ButtonGroup, { groupVariants, sizes } from "@/shared/components/ButtonGroup";
+import Card, { cardType } from "@/shared/components/Card";
+
+interface ToolConfirmationCardProps {
+  confirmation: ToolConfirmation;
+  onResolve: (confirmation: ToolConfirmation, decision: "approve" | "cancel") => void;
+}
+
+const ToolConfirmationCard = ({ confirmation, onResolve }: ToolConfirmationCardProps) => {
+  const isPending = confirmation.status === "pending";
+  const isSubmitting = confirmation.status === "submitting";
+
+  const buttons = [
+    {
+      text: "Cancel",
+      variant: variants.secondary,
+      disabled: isSubmitting,
+      onClick: () => onResolve(confirmation, "cancel"),
+    },
+    {
+      text: isSubmitting ? "Submitting" : confirmation.confirmLabel,
+      variant: variants.primary,
+      disabled: isSubmitting,
+      loading: isSubmitting,
+      onClick: () => onResolve(confirmation, "approve"),
+    },
+  ];
+
+  return (
+    <Card
+      bodyClassName="px-4 pb-4"
+      className="border border-border-10"
+      headerTone="neutral"
+      testId="tool-confirmation"
+      title={
+        <div className="flex min-w-0 items-center gap-3">
+          <span aria-hidden="true" className="shrink-0 text-text-warning">
+            <Alert />
+          </span>
+          <h3 id={`tool-confirmation-${confirmation.id}`} className="min-w-0 text-emphasis-300 text-text-primary">
+            {confirmation.title}
+          </h3>
+        </div>
+      }
+      type={cardType.default}
+    >
+      <section aria-labelledby={`tool-confirmation-${confirmation.id}`}>
+        <p className="text-200 text-text-primary-50">{confirmation.description}</p>
+
+        {confirmation.details.length > 0 ? (
+          <dl className="mt-3 divide-y divide-border-5 rounded-xl bg-core-primary-5 px-3">
+            {confirmation.details.map((detail) => (
+              <div
+                key={`${detail.label}-${detail.value}`}
+                className="grid grid-cols-[minmax(0,2fr)_minmax(0,3fr)] gap-3 py-2 text-200"
+              >
+                <dt className="text-text-primary-50">{detail.label}</dt>
+                <dd className="text-right break-words text-text-primary">{detail.value}</dd>
+              </div>
+            ))}
+          </dl>
+        ) : null}
+
+        {confirmation.error ? (
+          <p className="mt-3 text-200 text-text-critical" role="alert">
+            {confirmation.error}
+          </p>
+        ) : null}
+
+        {isPending || isSubmitting ? (
+          <ButtonGroup buttons={buttons} className="mt-4" size={sizes.compact} variant={groupVariants.rightAligned} />
+        ) : (
+          <p className="mt-3 text-emphasis-200 text-text-primary-50" role="status">
+            {confirmation.status === "approved" ? "Approved" : null}
+            {confirmation.status === "cancelled" ? "Cancelled" : null}
+            {confirmation.status === "expired" ? "Confirmation expired" : null}
+          </p>
+        )}
+      </section>
+    </Card>
+  );
+};
+
+export default ToolConfirmationCard;
diff --git a/client/src/protoFleet/features/aiChat/chatContext.test.ts b/client/src/protoFleet/features/aiChat/chatContext.test.ts
new file mode 100644
index 000000000..58e228b01
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/chatContext.test.ts
@@ -0,0 +1,59 @@
+import { describe, expect, test } from "vitest";
+
+import { getChatContext } from "./chatContext";
+
+const labelsFor = (pathname: string) => getChatContext(pathname).suggestions.map((suggestion) => suggestion.label);
+
+describe("chatContext", () => {
+  test("uses dashboard prompts as the default context", () => {
+    expect(labelsFor("/")).toEqual([
+      "Summarize fleet health",
+      "Find profit risks today",
+      "What needs attention first?",
+    ]);
+    expect(labelsFor("/dashboard")).toEqual([
+      "Summarize fleet health",
+      "Find profit risks today",
+      "What needs attention first?",
+    ]);
+  });
+
+  test("normalizes site-scoped routes before resolving context", () => {
+    expect(labelsFor("/cedar-creek/fleet/miners")).toEqual([
+      "Which miners are offline?",
+      "Find firmware drift",
+      "Check weak or shared passwords",
+    ]);
+    expect(labelsFor("/north-yard/energy")).toEqual([
+      "Summarize curtailment impact",
+      "Find profitable power changes",
+      "Plan today's power strategy",
+    ]);
+  });
+
+  test("uses broader prompts for the dedicated Minerbot page", () => {
+    expect(labelsFor("/minerbot")).toEqual([
+      "Start a fleet health review",
+      "Plan recurring work",
+      "Find tasks to automate",
+    ]);
+  });
+
+  test("uses more specific prompts for detail and settings routes", () => {
+    expect(labelsFor("/racks/12")).toEqual([
+      "Find rack hotspots",
+      "Show underperforming slots",
+      "Suggest miner placement fixes",
+    ]);
+    expect(labelsFor("/settings/agents")).toEqual([
+      "Check Minerbot setup",
+      "Help choose a provider",
+      "What can Minerbot do here?",
+    ]);
+    expect(labelsFor("/settings/firmware")).toEqual([
+      "Review firmware defaults",
+      "Find miners behind firmware",
+      "Plan a staged update",
+    ]);
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/chatContext.ts b/client/src/protoFleet/features/aiChat/chatContext.ts
new file mode 100644
index 000000000..beb694fcf
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/chatContext.ts
@@ -0,0 +1,217 @@
+import type { ChatSuggestion } from "./types";
+
+type ChatContextKey =
+  | "dashboard"
+  | "minerbot"
+  | "fleetMiners"
+  | "fleetRacks"
+  | "fleetSites"
+  | "fleetInfrastructure"
+  | "groups"
+  | "energy"
+  | "activity"
+  | "singleMiner"
+  | "settings"
+  | "settingsAgents"
+  | "settingsFirmware"
+  | "settingsSchedules"
+  | "settingsCurtailment"
+  | "onboarding";
+
+export interface ChatContext {
+  key: ChatContextKey;
+  description: string;
+  suggestions: ChatSuggestion[];
+}
+
+const chatContexts: Record<ChatContextKey, ChatContext> = {
+  dashboard: {
+    key: "dashboard",
+    description: "Ask about fleet health, profitability, and urgent risks.",
+    suggestions: [
+      { label: "Summarize fleet health" },
+      { label: "Find profit risks today" },
+      { label: "What needs attention first?" },
+    ],
+  },
+  minerbot: {
+    key: "minerbot",
+    description: "Ask Minerbot to review fleet signals, prepare actions, or plan recurring work.",
+    suggestions: [
+      { label: "Start a fleet health review" },
+      { label: "Plan recurring work" },
+      { label: "Find tasks to automate" },
+    ],
+  },
+  fleetMiners: {
+    key: "fleetMiners",
+    description: "Ask about miner status, firmware drift, and security risks.",
+    suggestions: [
+      { label: "Which miners are offline?" },
+      { label: "Find firmware drift" },
+      { label: "Check weak or shared passwords" },
+    ],
+  },
+  fleetRacks: {
+    key: "fleetRacks",
+    description: "Ask about rack health, placement, and underperforming slots.",
+    suggestions: [
+      { label: "Find rack hotspots" },
+      { label: "Show underperforming slots" },
+      { label: "Suggest miner placement fixes" },
+    ],
+  },
+  fleetSites: {
+    key: "fleetSites",
+    description: "Ask about site setup, building health, and rack configuration.",
+    suggestions: [
+      { label: "Review site setup" },
+      { label: "Find building health risks" },
+      { label: "Suggest rack configuration fixes" },
+    ],
+  },
+  fleetInfrastructure: {
+    key: "fleetInfrastructure",
+    description: "Ask about infrastructure health, network risks, and devices needing attention.",
+    suggestions: [
+      { label: "Summarize infrastructure health" },
+      { label: "Find network risks" },
+      { label: "Review devices needing attention" },
+    ],
+  },
+  groups: {
+    key: "groups",
+    description: "Ask about group performance, outliers, and group-level actions.",
+    suggestions: [
+      { label: "Compare group performance" },
+      { label: "Find underperforming groups" },
+      { label: "Suggest group-level actions" },
+    ],
+  },
+  energy: {
+    key: "energy",
+    description: "Ask about curtailment impact, power pricing, and market-aware operations.",
+    suggestions: [
+      { label: "Summarize curtailment impact" },
+      { label: "Find profitable power changes" },
+      { label: "Plan today's power strategy" },
+    ],
+  },
+  activity: {
+    key: "activity",
+    description: "Ask about recent changes, incomplete operations, and operator activity.",
+    suggestions: [
+      { label: "Summarize recent activity" },
+      { label: "Find operations needing review" },
+      { label: "What changed today?" },
+    ],
+  },
+  singleMiner: {
+    key: "singleMiner",
+    description: "Ask about this miner's health, hashrate, and safe next actions.",
+    suggestions: [
+      { label: "Summarize this miner" },
+      { label: "Diagnose hashrate drop" },
+      { label: "Review safe actions" },
+    ],
+  },
+  settings: {
+    key: "settings",
+    description: "Ask about setup gaps, operating defaults, and recurring tasks.",
+    suggestions: [
+      { label: "Review setup gaps" },
+      { label: "Suggest operational defaults" },
+      { label: "Show recurring tasks to automate" },
+    ],
+  },
+  settingsAgents: {
+    key: "settingsAgents",
+    description: "Ask about Minerbot setup, provider choices, and available capabilities.",
+    suggestions: [
+      { label: "Check Minerbot setup" },
+      { label: "Help choose a provider" },
+      { label: "What can Minerbot do here?" },
+    ],
+  },
+  settingsFirmware: {
+    key: "settingsFirmware",
+    description: "Ask about firmware defaults, version drift, and staged updates.",
+    suggestions: [
+      { label: "Review firmware defaults" },
+      { label: "Find miners behind firmware" },
+      { label: "Plan a staged update" },
+    ],
+  },
+  settingsSchedules: {
+    key: "settingsSchedules",
+    description: "Ask about schedules, conflicts, and recurring operating tasks.",
+    suggestions: [
+      { label: "Review power schedules" },
+      { label: "Find schedule conflicts" },
+      { label: "Suggest recurring tasks" },
+    ],
+  },
+  settingsCurtailment: {
+    key: "settingsCurtailment",
+    description: "Ask about curtailment settings, market windows, and power rules.",
+    suggestions: [
+      { label: "Review curtailment settings" },
+      { label: "Find profitable curtailment windows" },
+      { label: "Suggest market-based rules" },
+    ],
+  },
+  onboarding: {
+    key: "onboarding",
+    description: "Ask about site setup, miner configuration, and launch checks.",
+    suggestions: [
+      { label: "Plan site setup" },
+      { label: "Validate miner configuration" },
+      { label: "What should I check before launch?" },
+    ],
+  },
+};
+
+const SCOPABLE_ROUTE_SEGMENTS = new Set(["dashboard", "fleet", "groups", "energy", "activity"]);
+
+const pathSegments = (pathname: string) =>
+  pathname
+    .split("/")
+    .map((segment) => segment.trim())
+    .filter(Boolean);
+
+const normalizedSegments = (pathname: string) => {
+  const segments = pathSegments(pathname);
+  return segments.length > 1 && SCOPABLE_ROUTE_SEGMENTS.has(segments[1]) ? segments.slice(1) : segments;
+};
+
+export const getChatContext = (pathname: string): ChatContext => {
+  const [primary, secondary] = normalizedSegments(pathname);
+
+  if (!primary || primary === "dashboard") return chatContexts.dashboard;
+  if (primary === "minerbot") return chatContexts.minerbot;
+  if (primary === "groups") return chatContexts.groups;
+  if (primary === "energy") return chatContexts.energy;
+  if (primary === "activity") return chatContexts.activity;
+  if (primary === "miners") return chatContexts.singleMiner;
+  if (primary === "racks") return chatContexts.fleetRacks;
+  if (primary === "sites" || primary === "buildings") return chatContexts.fleetSites;
+  if (primary === "onboarding") return chatContexts.onboarding;
+
+  if (primary === "fleet") {
+    if (secondary === "miners") return chatContexts.fleetMiners;
+    if (secondary === "racks") return chatContexts.fleetRacks;
+    if (secondary === "sites" || secondary === "buildings") return chatContexts.fleetSites;
+    if (secondary === "infrastructure") return chatContexts.fleetInfrastructure;
+    return chatContexts.fleetMiners;
+  }
+
+  if (primary === "settings") {
+    if (secondary === "agents") return chatContexts.settingsAgents;
+    if (secondary === "firmware") return chatContexts.settingsFirmware;
+    if (secondary === "schedules") return chatContexts.settingsSchedules;
+    if (secondary === "curtailment") return chatContexts.settingsCurtailment;
+    return chatContexts.settings;
+  }
+
+  return chatContexts.dashboard;
+};
diff --git a/client/src/protoFleet/features/aiChat/index.ts b/client/src/protoFleet/features/aiChat/index.ts
new file mode 100644
index 000000000..d25ac50ed
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/index.ts
@@ -0,0 +1,4 @@
+export { default as ChatFab } from "./ChatFab";
+export { default as ChatPanel } from "./ChatPanel";
+export { useChatStore } from "./useChatStore";
+export type { LLMProviderConfig, ChatMessage, AgentActivity } from "./types";
diff --git a/client/src/protoFleet/features/aiChat/pages/MinerbotPage.test.tsx b/client/src/protoFleet/features/aiChat/pages/MinerbotPage.test.tsx
new file mode 100644
index 000000000..7fd96d1fa
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/pages/MinerbotPage.test.tsx
@@ -0,0 +1,165 @@
+import { MemoryRouter } from "react-router-dom";
+import { act, fireEvent, render, screen, waitFor, within } from "@testing-library/react";
+import { afterEach, beforeEach, describe, expect, test, vi } from "vitest";
+
+import MinerbotPage from "./MinerbotPage";
+import { useChatStore } from "@/protoFleet/features/aiChat/useChatStore";
+
+const mocks = vi.hoisted(() => ({
+  sendMessage: vi.fn(),
+  resolveToolConfirmation: vi.fn(),
+}));
+const scrollIntoViewMock = vi.fn();
+
+vi.mock("@/protoFleet/api/clients", () => ({
+  chatClient: mocks,
+}));
+
+describe("MinerbotPage", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    scrollIntoViewMock.mockClear();
+    Element.prototype.scrollIntoView = scrollIntoViewMock;
+    mocks.sendMessage.mockImplementation(() => (async function* () {})());
+    act(() => {
+      useChatStore.getState().clearMessages();
+    });
+  });
+
+  afterEach(() => {
+    act(() => {
+      useChatStore.getState().clearMessages();
+    });
+  });
+
+  test("renders the cleaned-up chat surface with actionable suggestion cards and history", () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <MinerbotPage />
+      </MemoryRouter>,
+    );
+
+    const history = screen.getByLabelText("Chat history");
+    const historyRows = within(history).getAllByRole("listitem");
+    const suggestions = screen.getByLabelText("Actionable suggestions");
+
+    expect(screen.queryByRole("heading", { name: "Minerbot" })).not.toBeInTheDocument();
+    expect(screen.queryByRole("heading", { name: "Chat" })).not.toBeInTheDocument();
+    expect(screen.getByRole("navigation", { name: "Minerbot" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "New chat" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "Suggestions" })).toHaveAttribute("aria-current", "page");
+    expect(historyRows[0]).toHaveTextContent("Firmware drift review");
+    expect(historyRows[0]).not.toHaveTextContent("2 hours ago");
+    expect(historyRows[1]).toHaveTextContent("Power strategy");
+    expect(historyRows[1]).not.toHaveTextContent("Yesterday");
+    expect(screen.queryByRole("heading", { name: "Suggested workflows" })).not.toBeInTheDocument();
+    expect(within(suggestions).getAllByTestId("minerbot-suggestion-card")).toHaveLength(6);
+    expect(screen.getByRole("heading", { name: "Forecast failing hardware" })).toBeInTheDocument();
+    expect(screen.getByRole("button", { name: "Forecast failures" })).toBeInTheDocument();
+  });
+
+  test("sends actionable suggestions through the shared Minerbot conversation path", async () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <MinerbotPage />
+      </MemoryRouter>,
+    );
+
+    fireEvent.click(screen.getByRole("button", { name: "Forecast failures" }));
+
+    await waitFor(() => {
+      expect(mocks.sendMessage).toHaveBeenCalledWith(
+        expect.objectContaining({
+          content: "Forecast failing hardware and recommend the highest-priority repairs.",
+        }),
+        expect.any(Object),
+      );
+    });
+    expect(
+      within(screen.getByLabelText("Conversation")).getByText(
+        "Forecast failing hardware and recommend the highest-priority repairs.",
+      ),
+    ).toBeInTheDocument();
+  });
+
+  test("loads previous chats from history", () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <MinerbotPage />
+      </MemoryRouter>,
+    );
+
+    fireEvent.click(screen.getByRole("button", { name: /Firmware drift review/ }));
+
+    const conversation = screen.getByLabelText("Conversation");
+    const scrollArea = screen.getByTestId("minerbot-chat-scroll-area");
+
+    expect(screen.getByRole("button", { name: /Firmware drift review/ })).toHaveAttribute("aria-current", "page");
+    expect(screen.queryByLabelText("Actionable suggestions")).not.toBeInTheDocument();
+    expect(scrollArea).toHaveClass("min-h-0", "flex-1", "overflow-y-auto", "scroll-pb-8");
+    expect(scrollIntoViewMock).toHaveBeenLastCalledWith({ behavior: "smooth", block: "end", inline: "nearest" });
+    expect(conversation).toHaveClass("max-w-[800px]");
+    expect(within(conversation).getByRole("heading", { name: "Firmware drift review" })).toBeInTheDocument();
+    expect(within(conversation).getByText("2 hours ago")).toBeInTheDocument();
+    expect(within(conversation).getByText("Find miners behind firmware and plan a staged update.")).toBeInTheDocument();
+    expect(within(conversation).getByText(/I found 8 miners behind/)).toBeInTheDocument();
+    expect(
+      within(conversation)
+        .getByText("Find miners behind firmware and plan a staged update.")
+        .closest(".bg-core-primary-fill"),
+    ).not.toBeNull();
+    expect(
+      within(conversation)
+        .getByText(/I found 8 miners behind/)
+        .closest(".bg-core-primary-5"),
+    ).toBeNull();
+    expect(screen.getByRole("table")).toBeInTheDocument();
+    expect(mocks.sendMessage).not.toHaveBeenCalled();
+  });
+
+  test("starts an empty chat from a loaded history thread", () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <MinerbotPage />
+      </MemoryRouter>,
+    );
+
+    fireEvent.click(screen.getByRole("button", { name: /Firmware drift review/ }));
+    fireEvent.click(screen.getByRole("button", { name: "New chat" }));
+
+    const conversation = screen.getByLabelText("Conversation");
+    const promptIdeas = screen.getByLabelText("Suggested prompts");
+
+    expect(screen.getByRole("button", { name: "New chat" })).toHaveAttribute("aria-current", "page");
+    expect(screen.queryByLabelText("Actionable suggestions")).not.toBeInTheDocument();
+    expect(screen.getByRole("heading", { name: "What would you like to know?" })).toBeInTheDocument();
+    expect(within(promptIdeas).getAllByRole("button")).toHaveLength(3);
+    expect(within(promptIdeas).getByRole("button", { name: "Start a fleet health review" })).toBeInTheDocument();
+    expect(
+      within(conversation).queryByText("Find miners behind firmware and plan a staged update."),
+    ).not.toBeInTheDocument();
+    expect(within(conversation).queryByText(/I found 8 miners behind/)).not.toBeInTheDocument();
+    expect(mocks.sendMessage).not.toHaveBeenCalled();
+  });
+
+  test("starts a new chat from a prompt idea", async () => {
+    render(
+      <MemoryRouter initialEntries={["/minerbot"]}>
+        <MinerbotPage />
+      </MemoryRouter>,
+    );
+
+    fireEvent.click(screen.getByRole("button", { name: "New chat" }));
+    fireEvent.click(screen.getByRole("button", { name: "Plan recurring work" }));
+
+    await waitFor(() => {
+      expect(mocks.sendMessage).toHaveBeenCalledWith(
+        expect.objectContaining({
+          content: "Plan recurring work",
+        }),
+        expect.any(Object),
+      );
+    });
+    expect(within(screen.getByLabelText("Conversation")).getByText("Plan recurring work")).toBeInTheDocument();
+  });
+});
diff --git a/client/src/protoFleet/features/aiChat/pages/MinerbotPage.tsx b/client/src/protoFleet/features/aiChat/pages/MinerbotPage.tsx
new file mode 100644
index 000000000..4b92b7f23
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/pages/MinerbotPage.tsx
@@ -0,0 +1,254 @@
+import { type ReactNode, useCallback, useEffect, useRef, useState } from "react";
+import clsx from "clsx";
+
+import ChatConversation from "../ChatConversation";
+import ChatInput from "../ChatInput";
+import useMinerbotConversation from "../useMinerbotConversation";
+import {
+  type MinerbotHistoryThread,
+  minerbotHistoryThreads,
+  minerbotSuggestionCards,
+  type MinerbotSuggestionIcon,
+} from "./minerbotExperience";
+import { Activity, Efficiency, Graph, type IconProps, LightningAlt, Lock, Settings } from "@/shared/assets/icons";
+import Button, { sizes as buttonSizes, variants as buttonVariants } from "@/shared/components/Button";
+
+const suggestionIconByKey: Record<MinerbotSuggestionIcon, (props: IconProps) => ReactNode> = {
+  activity: Activity,
+  energy: LightningAlt,
+  firmware: Settings,
+  onboarding: Graph,
+  profitability: Efficiency,
+  security: Lock,
+};
+
+const MinerbotPage = () => {
+  const {
+    chatContext,
+    conversationItems,
+    hasConversation,
+    hasPendingConfirmation,
+    isStreaming,
+    loadConversation,
+    openSettings,
+    resolveConfirmation,
+    sendMessage,
+    startNewChat,
+    streamError,
+    streamingContent,
+  } = useMinerbotConversation();
+  const [activeHistoryId, setActiveHistoryId] = useState<string | null>(null);
+  const [activeView, setActiveView] = useState<"chat" | "suggestions">(hasConversation ? "chat" : "suggestions");
+  const conversationEndRef = useRef<HTMLDivElement>(null);
+
+  useEffect(() => {
+    conversationEndRef.current?.scrollIntoView({ behavior: "smooth", block: "end", inline: "nearest" });
+  }, [conversationItems, isStreaming, streamError, streamingContent]);
+
+  const handleStartNewChat = useCallback(() => {
+    setActiveHistoryId(null);
+    setActiveView("chat");
+    startNewChat();
+  }, [startNewChat]);
+
+  const handleShowSuggestions = useCallback(() => {
+    setActiveHistoryId(null);
+    setActiveView("suggestions");
+    startNewChat();
+  }, [startNewChat]);
+
+  const handleSendMessage = useCallback(
+    (content: string) => {
+      setActiveHistoryId(null);
+      setActiveView("chat");
+      sendMessage(content);
+    },
+    [sendMessage],
+  );
+
+  const handleSuggestionClick = useCallback(
+    (prompt: string) => {
+      setActiveHistoryId(null);
+      setActiveView("chat");
+      sendMessage(prompt);
+    },
+    [sendMessage],
+  );
+
+  const handleLoadHistory = useCallback(
+    (thread: MinerbotHistoryThread) => {
+      setActiveHistoryId(thread.id);
+      setActiveView("chat");
+      loadConversation(thread.id, thread.messages);
+    },
+    [loadConversation],
+  );
+
+  const showSuggestions = activeView === "suggestions";
+  const showCurrentChat = activeView === "chat" && !activeHistoryId;
+  const showEmptyChatPrompts = showCurrentChat && !hasConversation;
+  const activeHistoryThread = minerbotHistoryThreads.find((thread) => thread.id === activeHistoryId);
+
+  const navItemClassName = (active = false) =>
+    clsx(
+      "block w-full rounded-lg px-2 py-1.5 text-left text-emphasis-300 transition-colors",
+      "hover:bg-core-primary-5 focus-visible:ring-2 focus-visible:ring-core-primary-fill focus-visible:ring-offset-2 focus-visible:ring-offset-surface-base focus-visible:outline-none",
+      active ? "bg-core-primary-5 text-text-primary" : "text-text-primary-70",
+    );
+
+  return (
+    <div className="grid h-full min-h-0 overflow-hidden text-text-primary laptop:grid-cols-[260px_minmax(0,1fr)]">
+      <aside className="min-h-0 overflow-y-auto border-b border-border-5 bg-surface-base laptop:border-r laptop:border-b-0">
+        <nav aria-label="Minerbot">
+          <ul className="flex w-full shrink-0 flex-col gap-8 px-3 pt-6 text-text-primary-70">
+            <li>
+              <ul className="flex flex-col">
+                <li>
+                  <button
+                    aria-current={showCurrentChat ? "page" : undefined}
+                    className={navItemClassName(showCurrentChat)}
+                    onClick={handleStartNewChat}
+                    type="button"
+                  >
+                    New chat
+                  </button>
+                </li>
+                <li>
+                  <button
+                    aria-current={showSuggestions ? "page" : undefined}
+                    className={navItemClassName(showSuggestions)}
+                    onClick={handleShowSuggestions}
+                    type="button"
+                  >
+                    Suggestions
+                  </button>
+                </li>
+              </ul>
+            </li>
+            <li>
+              <div className="px-2 pb-2 text-300 font-medium text-text-primary-50">History</div>
+              <ul aria-label="Chat history" className="flex flex-col">
+                {minerbotHistoryThreads.map((item) => (
+                  <li key={item.title}>
+                    <button
+                      aria-current={activeHistoryId === item.id ? "page" : undefined}
+                      className={navItemClassName(activeHistoryId === item.id)}
+                      onClick={() => handleLoadHistory(item)}
+                      type="button"
+                    >
+                      <span className="block truncate">{item.title}</span>
+                    </button>
+                  </li>
+                ))}
+              </ul>
+            </li>
+          </ul>
+        </nav>
+      </aside>
+
+      <section className="flex min-h-0 flex-col overflow-hidden bg-surface-elevated-base">
+        <div
+          className="min-h-0 flex-1 scroll-pb-8 overflow-y-auto overscroll-contain px-5 py-5 laptop:px-8 laptop:py-8"
+          data-testid="minerbot-chat-scroll-area"
+        >
+          {showSuggestions ? (
+            <div className="mx-auto flex min-h-full w-full max-w-6xl flex-col justify-center gap-6 py-8">
+              <div className="max-w-2xl">
+                <h1 className="text-heading-300">What should Minerbot work on?</h1>
+                <p className="mt-2 text-300 text-text-primary-50">{chatContext.description}</p>
+              </div>
+              <div aria-label="Actionable suggestions" className="grid gap-3 tablet:grid-cols-2 desktop:grid-cols-3">
+                {minerbotSuggestionCards.map((suggestion) => {
+                  const SuggestionIcon = suggestionIconByKey[suggestion.icon];
+
+                  return (
+                    <article
+                      key={suggestion.title}
+                      className="flex min-h-48 flex-col rounded-lg border border-border-5 bg-surface-base p-4 shadow-50"
+                      data-testid="minerbot-suggestion-card"
+                    >
+                      <div className="flex items-start justify-between gap-3">
+                        <div className="flex size-8 shrink-0 items-center justify-center rounded-lg bg-core-accent-10 text-text-emphasis">
+                          <SuggestionIcon width="w-4" />
+                        </div>
+                        <span className="rounded-full bg-core-primary-5 px-2 py-1 text-200 font-medium whitespace-nowrap text-text-primary-50">
+                          {suggestion.impact}
+                        </span>
+                      </div>
+
+                      <h2 className="mt-4 text-heading-100">{suggestion.title}</h2>
+                      <p className="mt-1 flex-1 text-300 text-text-primary-70">{suggestion.description}</p>
+
+                      <div className="mt-4">
+                        <Button
+                          disabled={isStreaming}
+                          onClick={() => handleSuggestionClick(suggestion.prompt)}
+                          size={buttonSizes.compact}
+                          text={suggestion.actionLabel}
+                          variant={buttonVariants.secondary}
+                        />
+                      </div>
+                    </article>
+                  );
+                })}
+              </div>
+            </div>
+          ) : showEmptyChatPrompts ? (
+            <div
+              aria-label="Conversation"
+              aria-live="polite"
+              className="mx-auto flex min-h-full w-full max-w-[800px] flex-col justify-center gap-6 py-8"
+            >
+              <div>
+                <h1 className="text-heading-200 text-text-primary">What would you like to know?</h1>
+                <p className="mt-1 text-300 text-text-primary-50">{chatContext.description}</p>
+              </div>
+
+              <div aria-label="Suggested prompts" className="flex flex-wrap gap-2">
+                {chatContext.suggestions.map((suggestion) => (
+                  <Button
+                    key={suggestion.label}
+                    disabled={isStreaming}
+                    onClick={() => handleSuggestionClick(suggestion.label)}
+                    size={buttonSizes.compact}
+                    text={suggestion.label}
+                    variant={buttonVariants.secondary}
+                  />
+                ))}
+              </div>
+              <div ref={conversationEndRef} />
+            </div>
+          ) : (
+            <ChatConversation
+              conversationEndRef={conversationEndRef}
+              conversationItems={conversationItems}
+              header={
+                activeHistoryThread ? (
+                  <div>
+                    <h1 className="text-heading-200 text-text-primary">{activeHistoryThread.title}</h1>
+                    <p className="mt-1 text-300 text-text-primary-50">{activeHistoryThread.timeLabel}</p>
+                  </div>
+                ) : undefined
+              }
+              hasPendingConfirmation={hasPendingConfirmation}
+              isStreaming={isStreaming}
+              layout="fullScreen"
+              onOpenSettings={openSettings}
+              onResolveConfirmation={resolveConfirmation}
+              streamError={streamError}
+              streamingContent={streamingContent}
+            />
+          )}
+        </div>
+
+        <div className="shrink-0 border-t border-border-5 bg-surface-elevated-base px-5 pt-4 pb-5 laptop:px-8">
+          <div className="mx-auto w-full max-w-[800px]">
+            <ChatInput disabled={isStreaming} onSend={handleSendMessage} />
+          </div>
+        </div>
+      </section>
+    </div>
+  );
+};
+
+export default MinerbotPage;
diff --git a/client/src/protoFleet/features/aiChat/pages/minerbotExperience.ts b/client/src/protoFleet/features/aiChat/pages/minerbotExperience.ts
new file mode 100644
index 000000000..7bbc79e35
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/pages/minerbotExperience.ts
@@ -0,0 +1,112 @@
+import type { ChatTranscriptTurn } from "../types";
+
+export type MinerbotSuggestionIcon = "activity" | "energy" | "firmware" | "onboarding" | "profitability" | "security";
+
+export type MinerbotSuggestionCard = {
+  actionLabel: string;
+  description: string;
+  icon: MinerbotSuggestionIcon;
+  impact: string;
+  prompt: string;
+  title: string;
+};
+
+export type MinerbotHistoryThread = {
+  id: string;
+  messages: ChatTranscriptTurn[];
+  timeLabel: string;
+  title: string;
+};
+
+export const minerbotSuggestionCards: MinerbotSuggestionCard[] = [
+  {
+    actionLabel: "Forecast failures",
+    description: "Review fan, hashboard, PSU, and temperature signals to identify miners likely to fail next.",
+    icon: "activity",
+    impact: "Reduce downtime",
+    prompt: "Forecast failing hardware and recommend the highest-priority repairs.",
+    title: "Forecast failing hardware",
+  },
+  {
+    actionLabel: "Tune energy",
+    description: "Balance hashrate against power price, curtailment windows, and site-level efficiency.",
+    icon: "energy",
+    impact: "Protect margin",
+    prompt: "Tune the fleet for optimal energy spend based on market conditions.",
+    title: "Tune energy spend",
+  },
+  {
+    actionLabel: "Plan updates",
+    description: "Find firmware drift, group compatible miners, and prepare a staged update plan.",
+    icon: "firmware",
+    impact: "Improve efficiency",
+    prompt: "Find miners behind firmware and plan a staged update.",
+    title: "Keep firmware current",
+  },
+  {
+    actionLabel: "Validate setup",
+    description: "Check site, building, rack, pool, and miner configuration before new hardware comes online.",
+    icon: "onboarding",
+    impact: "Save setup time",
+    prompt: "Validate site, building, rack, and miner configuration before launch.",
+    title: "Automate onboarding checks",
+  },
+  {
+    actionLabel: "Review passwords",
+    description: "Identify weak or shared credentials and prepare a rotation plan that limits downtime.",
+    icon: "security",
+    impact: "Reduce exposure",
+    prompt: "Check weak or shared passwords and recommend a rotation plan.",
+    title: "Find password risk",
+  },
+  {
+    actionLabel: "Build report",
+    description: "Summarize hashprice, downtime, power cost, pool performance, and site contribution.",
+    icon: "profitability",
+    impact: "Drive profitability",
+    prompt: "Generate a P&L report with the top operational levers.",
+    title: "Generate P&L report",
+  },
+];
+
+export const minerbotHistoryThreads: MinerbotHistoryThread[] = [
+  {
+    id: "firmware-drift-review",
+    timeLabel: "2 hours ago",
+    title: "Firmware drift review",
+    messages: [
+      { role: "user", content: "Find miners behind firmware and plan a staged update." },
+      {
+        role: "assistant",
+        content:
+          "I found 8 miners behind the recommended firmware. Start with Rack R03 because its stale share rate is 4.8% above fleet median.\n\n| Group | Miners | Risk | Recommended action |\n| --- | ---: | --- | --- |\n| Rack R03 | 5 | High stale shares | Stage update after current payout window |\n| Rack R07 | 3 | Efficiency drift | Queue after R03 is stable |\n\nI would update in two batches and keep a rollback window open for 30 minutes after each batch.",
+      },
+    ],
+  },
+  {
+    id: "power-strategy",
+    timeLabel: "Yesterday",
+    title: "Power strategy",
+    messages: [
+      { role: "user", content: "Tune the fleet for optimal energy spend based on market conditions." },
+      {
+        role: "assistant",
+        content:
+          "Austin can preserve margin by reducing draw during the next high-price window and returning full hashrate when prices normalize.\n\n| Site | Window | Suggested action | Expected effect |\n| --- | --- | --- | --- |\n| Austin | 2:00 PM-5:00 PM | Curtail lowest-efficiency miners | Avoid negative-margin hash |\n| Tulsa | 6:00 PM-8:00 PM | Hold current profile | Stable margin |\n\nBefore applying this, I would ask for operator approval because it changes power state.",
+      },
+    ],
+  },
+  {
+    id: "password-risk-review",
+    timeLabel: "Monday",
+    title: "Password risk review",
+    messages: [
+      { role: "user", content: "Check weak or shared passwords and recommend a rotation plan." },
+      {
+        role: "assistant",
+        content:
+          "I found 17 miners with shared credentials across 3 racks. Rotate Rack A7 first because it has the largest overlap with externally reachable network paths.\n\n| Scope | Miners | Issue | Next step |\n| --- | ---: | --- | --- |\n| Rack A7 | 9 | Shared password | Rotate during maintenance window |\n| Rack C2 | 5 | Weak password | Queue credential update |\n| Rack D4 | 3 | Default username | Review access policy |\n\nI can prepare the rotation set, but should request confirmation before changing credentials.",
+      },
+    ],
+  },
+];
diff --git a/client/src/protoFleet/features/aiChat/types.ts b/client/src/protoFleet/features/aiChat/types.ts
new file mode 100644
index 000000000..6bea6c750
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/types.ts
@@ -0,0 +1,57 @@
+/** Core types for the AI Chat feature */
+
+export type MessageRole = "user" | "assistant";
+
+export interface ChatMessage {
+  id: string;
+  role: MessageRole;
+  content: string;
+  timestamp: Date;
+  sequence: number;
+}
+
+export type ChatTranscriptTurn = Pick<ChatMessage, "role" | "content">;
+
+export interface AgentActivity {
+  id: string;
+  summary: string;
+  status: "running" | "completed" | "failed" | "cancelled";
+  timestamp: Date;
+  sequence: number;
+}
+
+export interface ToolConfirmationDetail {
+  label: string;
+  value: string;
+}
+
+export interface ToolConfirmation {
+  id: string;
+  toolCallId: string;
+  title: string;
+  description: string;
+  confirmLabel: string;
+  details: ToolConfirmationDetail[];
+  status: "pending" | "submitting" | "approved" | "cancelled" | "expired";
+  decision?: "approve" | "cancel";
+  error?: string;
+  sequence: number;
+}
+
+export interface ChatSuggestion {
+  label: string;
+  icon?: "star" | "default";
+}
+
+/**
+ * Configuration for the LLM provider. This is a BYOLLM solution — operators
+ * configure their own provider.
+ */
+export interface LLMProviderConfig {
+  harness: "native" | "goose";
+  provider: "openai" | "anthropic" | "ollama";
+  apiKey?: string;
+  baseUrl?: string;
+  model?: string;
+  gooseBaseUrl?: string;
+}
diff --git a/client/src/protoFleet/features/aiChat/useChatStore.ts b/client/src/protoFleet/features/aiChat/useChatStore.ts
new file mode 100644
index 000000000..5050f47b1
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/useChatStore.ts
@@ -0,0 +1,187 @@
+import { create } from "zustand";
+import { immer } from "zustand/middleware/immer";
+
+import type { AgentActivity, ChatMessage, ChatSuggestion, ChatTranscriptTurn, ToolConfirmation } from "./types";
+
+interface ChatState {
+  isOpen: boolean;
+  messages: ChatMessage[];
+  agentActivities: AgentActivity[];
+  toolConfirmations: ToolConfirmation[];
+  isStreaming: boolean;
+  streamingContent: string;
+  streamError: string;
+  suggestions: ChatSuggestion[];
+  nextSequence: number;
+
+  toggle: () => void;
+  open: () => void;
+  close: () => void;
+  addMessage: (role: ChatMessage["role"], content: string) => void;
+  loadMessages: (turns: ChatTranscriptTurn[]) => void;
+  setStreaming: (streaming: boolean) => void;
+  appendStreamingContent: (content: string) => void;
+  setStreamError: (error: string) => void;
+  beginToolActivity: (id: string, summary: string) => void;
+  finishToolActivity: (id: string, success: boolean, summary: string, cancelled?: boolean) => void;
+  addToolConfirmation: (confirmation: Omit<ToolConfirmation, "status" | "sequence">) => void;
+  submitToolConfirmation: (id: string, decision: "approve" | "cancel") => void;
+  resolveToolConfirmation: (id: string, decision: "approve" | "cancel") => void;
+  failToolConfirmation: (id: string, error: string) => void;
+  expirePendingConfirmations: () => void;
+  resetStream: () => void;
+  clearMessages: () => void;
+}
+
+const DEFAULT_SUGGESTIONS: ChatSuggestion[] = [
+  { label: "Summarize fleet health", icon: "star" },
+  { label: "How many miners are offline?" },
+  { label: "Compare miner states by site" },
+  { label: "List my sites" },
+  { label: "Show configured mining pools" },
+];
+
+export const useChatStore = create<ChatState>()(
+  immer((set) => ({
+    isOpen: false,
+    messages: [],
+    agentActivities: [],
+    toolConfirmations: [],
+    isStreaming: false,
+    streamingContent: "",
+    streamError: "",
+    suggestions: DEFAULT_SUGGESTIONS,
+    nextSequence: 0,
+
+    toggle: () =>
+      set((state) => {
+        state.isOpen = !state.isOpen;
+      }),
+    open: () =>
+      set((state) => {
+        state.isOpen = true;
+      }),
+    close: () =>
+      set((state) => {
+        state.isOpen = false;
+      }),
+
+    addMessage: (role, content) =>
+      set((state) => {
+        state.messages.push({
+          id: crypto.randomUUID(),
+          role,
+          content,
+          timestamp: new Date(),
+          sequence: state.nextSequence,
+        });
+        state.nextSequence += 1;
+      }),
+    loadMessages: (turns) =>
+      set((state) => {
+        state.messages = turns.map((turn, index) => ({
+          id: crypto.randomUUID(),
+          role: turn.role,
+          content: turn.content,
+          timestamp: new Date(),
+          sequence: index,
+        }));
+        state.agentActivities = [];
+        state.toolConfirmations = [];
+        state.streamingContent = "";
+        state.streamError = "";
+        state.nextSequence = turns.length;
+      }),
+
+    setStreaming: (streaming) =>
+      set((state) => {
+        state.isStreaming = streaming;
+      }),
+    appendStreamingContent: (content) =>
+      set((state) => {
+        state.streamingContent += content;
+      }),
+    setStreamError: (error) =>
+      set((state) => {
+        state.streamError = error;
+      }),
+    beginToolActivity: (id, summary) =>
+      set((state) => {
+        state.agentActivities.push({
+          id,
+          summary,
+          status: "running",
+          timestamp: new Date(),
+          sequence: state.nextSequence,
+        });
+        state.nextSequence += 1;
+      }),
+    finishToolActivity: (id, success, summary, cancelled = false) =>
+      set((state) => {
+        const activity = state.agentActivities.find((candidate) => candidate.id === id);
+        if (!activity) return;
+        activity.summary = summary;
+        activity.status = cancelled ? "cancelled" : success ? "completed" : "failed";
+        const confirmation = state.toolConfirmations.find((candidate) => candidate.toolCallId === id);
+        if (confirmation?.decision) {
+          confirmation.status = confirmation.decision === "approve" ? "approved" : "cancelled";
+          confirmation.error = undefined;
+        }
+      }),
+    addToolConfirmation: (confirmation) =>
+      set((state) => {
+        state.toolConfirmations.push({
+          ...confirmation,
+          status: "pending",
+          sequence: state.nextSequence,
+        });
+        state.nextSequence += 1;
+      }),
+    submitToolConfirmation: (id, decision) =>
+      set((state) => {
+        const confirmation = state.toolConfirmations.find((candidate) => candidate.id === id);
+        if (!confirmation || confirmation.status !== "pending") return;
+        confirmation.status = "submitting";
+        confirmation.decision = decision;
+        confirmation.error = undefined;
+      }),
+    resolveToolConfirmation: (id, decision) =>
+      set((state) => {
+        const confirmation = state.toolConfirmations.find((candidate) => candidate.id === id);
+        if (!confirmation) return;
+        confirmation.status = decision === "approve" ? "approved" : "cancelled";
+        confirmation.decision = decision;
+        confirmation.error = undefined;
+      }),
+    failToolConfirmation: (id, error) =>
+      set((state) => {
+        const confirmation = state.toolConfirmations.find((candidate) => candidate.id === id);
+        if (!confirmation) return;
+        confirmation.status = "pending";
+        confirmation.error = error;
+      }),
+    expirePendingConfirmations: () =>
+      set((state) => {
+        state.toolConfirmations.forEach((confirmation) => {
+          if (confirmation.status === "pending" || confirmation.status === "submitting") {
+            confirmation.status = "expired";
+            confirmation.error = undefined;
+          }
+        });
+      }),
+    resetStream: () =>
+      set((state) => {
+        state.streamingContent = "";
+        state.streamError = "";
+      }),
+    clearMessages: () =>
+      set((state) => {
+        state.messages = [];
+        state.agentActivities = [];
+        state.toolConfirmations = [];
+        state.streamingContent = "";
+        state.streamError = "";
+        state.nextSequence = 0;
+      }),
+  })),
+);
diff --git a/client/src/protoFleet/features/aiChat/useMinerbotConversation.ts b/client/src/protoFleet/features/aiChat/useMinerbotConversation.ts
new file mode 100644
index 000000000..019a05c90
--- /dev/null
+++ b/client/src/protoFleet/features/aiChat/useMinerbotConversation.ts
@@ -0,0 +1,220 @@
+import { useCallback, useEffect, useMemo, useRef } from "react";
+import { useLocation, useNavigate } from "react-router-dom";
+
+import { getChatContext } from "./chatContext";
+import type { AgentActivity, ChatMessage, ChatTranscriptTurn, ToolConfirmation } from "./types";
+import { useChatStore } from "./useChatStore";
+import { chatClient } from "@/protoFleet/api/clients";
+import { ChatRole, ToolConfirmationDecision } from "@/protoFleet/api/generated/chat/v1/chat_pb";
+import { getErrorMessage } from "@/protoFleet/api/getErrorMessage";
+
+export type MinerbotConversationItem =
+  | { kind: "message"; sequence: number; message: ChatMessage }
+  | { kind: "activity"; sequence: number; activity: AgentActivity }
+  | { kind: "confirmation"; sequence: number; confirmation: ToolConfirmation };
+
+const useMinerbotConversation = () => {
+  const messages = useChatStore((state) => state.messages);
+  const agentActivities = useChatStore((state) => state.agentActivities);
+  const toolConfirmations = useChatStore((state) => state.toolConfirmations);
+  const isStreaming = useChatStore((state) => state.isStreaming);
+  const streamingContent = useChatStore((state) => state.streamingContent);
+  const streamError = useChatStore((state) => state.streamError);
+  const addMessage = useChatStore((state) => state.addMessage);
+  const loadMessages = useChatStore((state) => state.loadMessages);
+  const setStreaming = useChatStore((state) => state.setStreaming);
+  const appendStreamingContent = useChatStore((state) => state.appendStreamingContent);
+  const setStreamError = useChatStore((state) => state.setStreamError);
+  const beginToolActivity = useChatStore((state) => state.beginToolActivity);
+  const finishToolActivity = useChatStore((state) => state.finishToolActivity);
+  const addToolConfirmation = useChatStore((state) => state.addToolConfirmation);
+  const submitToolConfirmation = useChatStore((state) => state.submitToolConfirmation);
+  const resolveToolConfirmation = useChatStore((state) => state.resolveToolConfirmation);
+  const failToolConfirmation = useChatStore((state) => state.failToolConfirmation);
+  const expirePendingConfirmations = useChatStore((state) => state.expirePendingConfirmations);
+  const resetStream = useChatStore((state) => state.resetStream);
+  const clearMessages = useChatStore((state) => state.clearMessages);
+  const location = useLocation();
+  const navigate = useNavigate();
+  const chatContext = useMemo(() => getChatContext(location.pathname), [location.pathname]);
+  const conversationIdRef = useRef<string>(crypto.randomUUID());
+  const requestGenerationRef = useRef(0);
+  const activeRequestRef = useRef<AbortController | null>(null);
+  const hasConversation = messages.length > 0;
+  const hasPendingConfirmation = toolConfirmations.some(
+    (confirmation) => confirmation.status === "pending" || confirmation.status === "submitting",
+  );
+  const conversationItems = useMemo(
+    () =>
+      [
+        ...messages.map((message) => ({ kind: "message" as const, sequence: message.sequence, message })),
+        ...agentActivities.map((activity) => ({ kind: "activity" as const, sequence: activity.sequence, activity })),
+        ...toolConfirmations.map((confirmation) => ({
+          kind: "confirmation" as const,
+          sequence: confirmation.sequence,
+          confirmation,
+        })),
+      ].sort((first, second) => first.sequence - second.sequence),
+    [agentActivities, messages, toolConfirmations],
+  );
+
+  const resolveConfirmation = useCallback(
+    async (confirmation: ToolConfirmation, decision: "approve" | "cancel") => {
+      const requestGeneration = requestGenerationRef.current;
+      submitToolConfirmation(confirmation.id, decision);
+      try {
+        await chatClient.resolveToolConfirmation({
+          confirmationId: confirmation.id,
+          decision: decision === "approve" ? ToolConfirmationDecision.APPROVE : ToolConfirmationDecision.CANCEL,
+        });
+        if (requestGeneration === requestGenerationRef.current) {
+          resolveToolConfirmation(confirmation.id, decision);
+        }
+      } catch (error) {
+        if (requestGeneration === requestGenerationRef.current) {
+          failToolConfirmation(
+            confirmation.id,
+            getErrorMessage(error, "Minerbot could not submit this confirmation. Try again."),
+          );
+        }
+      }
+    },
+    [failToolConfirmation, resolveToolConfirmation, submitToolConfirmation],
+  );
+
+  const sendMessage = useCallback(
+    async (content: string) => {
+      activeRequestRef.current?.abort();
+      const requestGeneration = ++requestGenerationRef.current;
+      const abortController = new AbortController();
+      activeRequestRef.current = abortController;
+      addMessage("user", content);
+      resetStream();
+      setStreaming(true);
+
+      let assistantContent = "";
+      try {
+        const stream = chatClient.sendMessage(
+          {
+            conversationId: conversationIdRef.current,
+            content,
+            history: messages.map((message) => ({
+              role: message.role === "user" ? ChatRole.USER : ChatRole.ASSISTANT,
+              content: message.content,
+            })),
+          },
+          { signal: abortController.signal },
+        );
+        for await (const response of stream) {
+          if (requestGeneration !== requestGenerationRef.current) break;
+          switch (response.event.case) {
+            case "textDelta": {
+              const delta = response.event.value.content;
+              assistantContent += delta;
+              appendStreamingContent(delta);
+              break;
+            }
+            case "toolCall":
+              beginToolActivity(response.event.value.id, response.event.value.summary);
+              break;
+            case "toolResult":
+              finishToolActivity(
+                response.event.value.id,
+                response.event.value.success,
+                response.event.value.summary,
+                response.event.value.cancelled,
+              );
+              break;
+            case "confirmationRequired":
+              addToolConfirmation({
+                id: response.event.value.confirmationId,
+                toolCallId: response.event.value.toolCallId,
+                title: response.event.value.title,
+                description: response.event.value.description,
+                confirmLabel: response.event.value.confirmLabel,
+                details: response.event.value.details.map((detail) => ({ label: detail.label, value: detail.value })),
+              });
+              break;
+          }
+        }
+        if (requestGeneration === requestGenerationRef.current && assistantContent.trim()) {
+          addMessage("assistant", assistantContent);
+        }
+      } catch (error) {
+        if (requestGeneration === requestGenerationRef.current) {
+          expirePendingConfirmations();
+          setStreamError(getErrorMessage(error, "Minerbot could not complete this request."));
+        }
+      } finally {
+        if (requestGeneration === requestGenerationRef.current) {
+          activeRequestRef.current = null;
+          setStreaming(false);
+        }
+      }
+    },
+    [
+      addMessage,
+      addToolConfirmation,
+      appendStreamingContent,
+      beginToolActivity,
+      expirePendingConfirmations,
+      finishToolActivity,
+      messages,
+      resetStream,
+      setStreamError,
+      setStreaming,
+    ],
+  );
+
+  const startNewChat = useCallback(() => {
+    requestGenerationRef.current += 1;
+    activeRequestRef.current?.abort();
+    activeRequestRef.current = null;
+    setStreaming(false);
+    conversationIdRef.current = crypto.randomUUID();
+    clearMessages();
+  }, [clearMessages, setStreaming]);
+
+  const loadConversation = useCallback(
+    (conversationId: string, turns: ChatTranscriptTurn[]) => {
+      requestGenerationRef.current += 1;
+      activeRequestRef.current?.abort();
+      activeRequestRef.current = null;
+      setStreaming(false);
+      conversationIdRef.current = conversationId;
+      loadMessages(turns);
+    },
+    [loadMessages, setStreaming],
+  );
+
+  const openSettings = useCallback(() => {
+    navigate("/settings/agents");
+  }, [navigate]);
+
+  useEffect(
+    () => () => {
+      requestGenerationRef.current += 1;
+      activeRequestRef.current?.abort();
+      activeRequestRef.current = null;
+      setStreaming(false);
+    },
+    [setStreaming],
+  );
+
+  return {
+    chatContext,
+    conversationItems,
+    hasConversation,
+    hasPendingConfirmation,
+    isStreaming,
+    loadConversation,
+    openSettings,
+    resolveConfirmation,
+    sendMessage,
+    startNewChat,
+    streamError,
+    streamingContent,
+  };
+};
+
+export default useMinerbotConversation;
diff --git a/client/src/protoFleet/features/settings/agents/Agents.tsx b/client/src/protoFleet/features/settings/agents/Agents.tsx
new file mode 100644
index 000000000..0aa87d24b
--- /dev/null
+++ b/client/src/protoFleet/features/settings/agents/Agents.tsx
@@ -0,0 +1,11 @@
+import LLMProviderSettings from "@/protoFleet/features/settings/agents/LLMProviderSettings";
+import SettingsPageHeader from "@/protoFleet/features/settings/components/SettingsPageHeader";
+
+const Agents = () => (
+  <div className="flex flex-col gap-8">
+    <SettingsPageHeader title="Agents" description="Configure the agent harness and model provider used by Minerbot." />
+    <LLMProviderSettings />
+  </div>
+);
+
+export default Agents;
diff --git a/client/src/protoFleet/features/settings/agents/LLMProviderSettings.test.tsx b/client/src/protoFleet/features/settings/agents/LLMProviderSettings.test.tsx
new file mode 100644
index 000000000..008c9e7d1
--- /dev/null
+++ b/client/src/protoFleet/features/settings/agents/LLMProviderSettings.test.tsx
@@ -0,0 +1,172 @@
+import { fireEvent, render, screen, waitFor } from "@testing-library/react";
+import { beforeEach, describe, expect, test, vi } from "vitest";
+
+import { AgentHarness, LLMProvider } from "@/protoFleet/api/generated/chat/v1/chat_pb";
+import LLMProviderSettings from "@/protoFleet/features/settings/agents/LLMProviderSettings";
+
+const mocks = vi.hoisted(() => ({
+  discoverModels: vi.fn(),
+  getLLMConfig: vi.fn(),
+  updateLLMConfig: vi.fn(),
+}));
+
+vi.mock("@/protoFleet/api/clients", () => ({
+  chatClient: mocks,
+}));
+
+describe("LLMProviderSettings", () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    mocks.discoverModels.mockResolvedValue({ models: [] });
+    mocks.getLLMConfig.mockResolvedValue({
+      config: {
+        harness: AgentHarness.NATIVE,
+        provider: LLMProvider.LLM_PROVIDER_UNSPECIFIED,
+        hasApiKey: false,
+        baseUrl: "",
+        model: "",
+        gooseBaseUrl: "",
+        hasGooseSecret: false,
+        configured: false,
+      },
+    });
+  });
+
+  test("loads remote-provider models after the operator enters an API key", async () => {
+    mocks.discoverModels.mockResolvedValue({
+      models: [{ id: "gpt-test", displayName: "GPT Test" }],
+    });
+    render(<LLMProviderSettings />);
+
+    const providerSelect = await screen.findByRole("button", { name: "Model provider" });
+    expect(providerSelect).toHaveTextContent("Select a provider");
+
+    fireEvent.click(providerSelect);
+    expect(screen.queryByText("Custom OpenAI-compatible")).not.toBeInTheDocument();
+    fireEvent.click(screen.getByText("OpenAI"));
+
+    fireEvent.change(screen.getByLabelText("API key"), { target: { value: "test-key" } });
+
+    await waitFor(() =>
+      expect(mocks.discoverModels).toHaveBeenCalledWith({
+        provider: LLMProvider.LLM_PROVIDER_OPENAI,
+        apiKey: "test-key",
+        baseUrl: "https://api.openai.com/v1",
+        useStoredApiKey: false,
+      }),
+    );
+
+    expect(screen.queryByRole("button", { name: /fetch models/i })).not.toBeInTheDocument();
+    const modelSelect = screen.getByRole("button", { name: "Model" });
+    expect(modelSelect).toHaveTextContent("Select a model");
+    expect(screen.getByRole("button", { name: "Save Minerbot settings" })).toBeDisabled();
+
+    fireEvent.click(modelSelect);
+    fireEvent.click(screen.getByText("GPT Test (gpt-test)"));
+    expect(screen.getByRole("button", { name: "Save Minerbot settings" })).toBeEnabled();
+  });
+
+  test("loads the saved provider models on page load", async () => {
+    mocks.getLLMConfig.mockResolvedValue({
+      config: {
+        harness: AgentHarness.NATIVE,
+        provider: LLMProvider.LLM_PROVIDER_ANTHROPIC,
+        hasApiKey: true,
+        baseUrl: "https://api.anthropic.com",
+        model: "claude-saved",
+        gooseBaseUrl: "",
+        hasGooseSecret: false,
+        configured: true,
+      },
+    });
+    mocks.discoverModels.mockResolvedValue({
+      models: [
+        { id: "claude-new", displayName: "Claude New" },
+        { id: "claude-saved", displayName: "Claude Saved" },
+      ],
+    });
+
+    render(<LLMProviderSettings />);
+
+    await waitFor(() =>
+      expect(mocks.discoverModels).toHaveBeenCalledWith({
+        provider: LLMProvider.LLM_PROVIDER_ANTHROPIC,
+        apiKey: "",
+        baseUrl: "https://api.anthropic.com",
+        useStoredApiKey: true,
+      }),
+    );
+    expect(screen.getByRole("button", { name: "Model" })).toHaveTextContent("Claude Saved");
+  });
+
+  test("loads Ollama models when the provider is selected", async () => {
+    mocks.discoverModels.mockResolvedValue({
+      models: [{ id: "llama-test:latest", displayName: "llama-test:latest" }],
+    });
+    render(<LLMProviderSettings />);
+
+    const providerSelect = await screen.findByRole("button", { name: "Model provider" });
+    fireEvent.click(providerSelect);
+    fireEvent.click(screen.getByText("Ollama"));
+
+    await waitFor(() =>
+      expect(mocks.discoverModels).toHaveBeenCalledWith({
+        provider: LLMProvider.LLM_PROVIDER_OLLAMA,
+        apiKey: "",
+        baseUrl: "http://127.0.0.1:11434",
+        useStoredApiKey: false,
+      }),
+    );
+    expect(screen.getByRole("button", { name: "Model" })).toHaveTextContent("Select a model");
+  });
+
+  test("can remove a stored API key without losing the selected model", async () => {
+    mocks.getLLMConfig.mockResolvedValue({
+      config: {
+        harness: AgentHarness.NATIVE,
+        provider: LLMProvider.LLM_PROVIDER_OPENAI,
+        hasApiKey: true,
+        baseUrl: "https://api.openai.com/v1",
+        model: "gpt-saved",
+        gooseBaseUrl: "",
+        hasGooseSecret: false,
+        configured: true,
+      },
+    });
+    mocks.discoverModels.mockResolvedValue({
+      models: [{ id: "gpt-saved", displayName: "GPT Saved" }],
+    });
+    mocks.updateLLMConfig.mockResolvedValue({
+      config: {
+        harness: AgentHarness.NATIVE,
+        provider: LLMProvider.LLM_PROVIDER_OPENAI,
+        hasApiKey: false,
+        baseUrl: "https://api.openai.com/v1",
+        model: "gpt-saved",
+        gooseBaseUrl: "",
+        hasGooseSecret: false,
+        configured: false,
+      },
+    });
+
+    render(<LLMProviderSettings />);
+
+    const removeButton = await screen.findByRole("button", { name: "Remove key" });
+    await waitFor(() => expect(screen.getByRole("button", { name: "Model" })).toHaveTextContent("GPT Saved"));
+    fireEvent.click(removeButton);
+
+    const saveButton = screen.getByRole("button", { name: "Save Minerbot settings" });
+    expect(saveButton).toBeEnabled();
+    fireEvent.click(saveButton);
+
+    await waitFor(() =>
+      expect(mocks.updateLLMConfig).toHaveBeenCalledWith(
+        expect.objectContaining({
+          apiKey: "",
+          clearApiKey: true,
+          model: "gpt-saved",
+        }),
+      ),
+    );
+  });
+});
diff --git a/client/src/protoFleet/features/settings/agents/LLMProviderSettings.tsx b/client/src/protoFleet/features/settings/agents/LLMProviderSettings.tsx
new file mode 100644
index 000000000..d8c1f8953
--- /dev/null
+++ b/client/src/protoFleet/features/settings/agents/LLMProviderSettings.tsx
@@ -0,0 +1,429 @@
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
+
+import { chatClient } from "@/protoFleet/api/clients";
+import { AgentHarness, type LLMConfig, LLMProvider } from "@/protoFleet/api/generated/chat/v1/chat_pb";
+import { getErrorMessage } from "@/protoFleet/api/getErrorMessage";
+import Button, { sizes, variants } from "@/shared/components/Button";
+import Input from "@/shared/components/Input";
+import Select from "@/shared/components/Select";
+import { pushToast, STATUSES } from "@/shared/features/toaster";
+
+type FormState = {
+  harness: AgentHarness;
+  provider: LLMProvider;
+  apiKey: string;
+  baseUrl: string;
+  model: string;
+  gooseBaseUrl: string;
+  gooseSecret: string;
+  hasApiKey: boolean;
+  hasGooseSecret: boolean;
+  clearApiKey: boolean;
+  clearGooseSecret: boolean;
+  storedApiKeyProvider: LLMProvider;
+};
+
+type DiscoveredModel = {
+  id: string;
+  displayName: string;
+};
+
+const EMPTY_FORM: FormState = {
+  harness: AgentHarness.NATIVE,
+  provider: LLMProvider.LLM_PROVIDER_UNSPECIFIED,
+  apiKey: "",
+  baseUrl: "",
+  model: "",
+  gooseBaseUrl: "",
+  gooseSecret: "",
+  hasApiKey: false,
+  hasGooseSecret: false,
+  clearApiKey: false,
+  clearGooseSecret: false,
+  storedApiKeyProvider: LLMProvider.LLM_PROVIDER_UNSPECIFIED,
+};
+
+const HARNESS_OPTIONS = [
+  {
+    value: String(AgentHarness.NATIVE),
+    label: "Embedded agent",
+    description: "Runs the provider and read-only fleet tool loop inside Fleet.",
+  },
+  {
+    value: String(AgentHarness.GOOSE),
+    label: "Goose ACP",
+    description: "Saves remote ACP settings; the server adapter is not enabled yet.",
+  },
+];
+
+const PROVIDER_OPTIONS = [
+  { value: String(LLMProvider.LLM_PROVIDER_OPENAI), label: "OpenAI" },
+  { value: String(LLMProvider.LLM_PROVIDER_ANTHROPIC), label: "Anthropic" },
+  { value: String(LLMProvider.LLM_PROVIDER_OLLAMA), label: "Ollama" },
+];
+
+const DEFAULT_BASE_URLS: Partial<Record<LLMProvider, string>> = {
+  [LLMProvider.LLM_PROVIDER_OPENAI]: "https://api.openai.com/v1",
+  [LLMProvider.LLM_PROVIDER_ANTHROPIC]: "https://api.anthropic.com",
+  [LLMProvider.LLM_PROVIDER_OLLAMA]: "http://127.0.0.1:11434",
+};
+
+const isSupportedProvider = (provider: LLMProvider) =>
+  provider === LLMProvider.LLM_PROVIDER_OPENAI ||
+  provider === LLMProvider.LLM_PROVIDER_ANTHROPIC ||
+  provider === LLMProvider.LLM_PROVIDER_OLLAMA;
+
+const formFromConfig = (config?: LLMConfig): FormState => {
+  const configuredProvider = config?.provider ?? LLMProvider.LLM_PROVIDER_UNSPECIFIED;
+  const provider = isSupportedProvider(configuredProvider) ? configuredProvider : LLMProvider.LLM_PROVIDER_UNSPECIFIED;
+  return {
+    ...EMPTY_FORM,
+    harness: config?.harness || AgentHarness.NATIVE,
+    provider,
+    baseUrl:
+      provider === configuredProvider
+        ? config?.baseUrl || DEFAULT_BASE_URLS[provider] || ""
+        : DEFAULT_BASE_URLS[provider] || "",
+    model: provider === configuredProvider ? (config?.model ?? "") : "",
+    gooseBaseUrl: config?.gooseBaseUrl ?? "",
+    hasApiKey: provider === configuredProvider && (config?.hasApiKey ?? false),
+    hasGooseSecret: config?.hasGooseSecret ?? false,
+    storedApiKeyProvider:
+      provider === configuredProvider && config?.hasApiKey ? provider : LLMProvider.LLM_PROVIDER_UNSPECIFIED,
+  };
+};
+
+const modelLabel = (model: DiscoveredModel) =>
+  model.displayName && model.displayName !== model.id ? `${model.displayName} (${model.id})` : model.id;
+
+const MODEL_DISCOVERY_DEBOUNCE_MS = 300;
+
+const LLMProviderSettings = () => {
+  const [form, setForm] = useState<FormState>(EMPTY_FORM);
+  const [isLoading, setIsLoading] = useState(true);
+  const [isSaving, setIsSaving] = useState(false);
+  const [isDiscovering, setIsDiscovering] = useState(false);
+  const [loadError, setLoadError] = useState("");
+  const [modelError, setModelError] = useState("");
+  const [models, setModels] = useState<DiscoveredModel[]>([]);
+  const discoveryRequestRef = useRef(0);
+
+  const loadConfig = useCallback(async () => {
+    setIsLoading(true);
+    setLoadError("");
+    try {
+      const response = await chatClient.getLLMConfig({});
+      const nextForm = formFromConfig(response.config);
+      setForm(nextForm);
+      setModels(nextForm.model ? [{ id: nextForm.model, displayName: nextForm.model }] : []);
+    } catch (error) {
+      setLoadError(getErrorMessage(error, "Could not load Minerbot settings."));
+    } finally {
+      setIsLoading(false);
+    }
+  }, []);
+
+  useEffect(() => {
+    // eslint-disable-next-line react-hooks/set-state-in-effect -- initial async server sync on settings mount
+    void loadConfig();
+  }, [loadConfig]);
+
+  const hasProvider = form.provider !== LLMProvider.LLM_PROVIDER_UNSPECIFIED;
+  const requiresAPIKey = hasProvider && form.provider !== LLMProvider.LLM_PROVIDER_OLLAMA;
+  const isGoose = form.harness === AgentHarness.GOOSE;
+  const canSave = useMemo(
+    () =>
+      !isLoading &&
+      !isSaving &&
+      hasProvider &&
+      form.model.trim().length > 0 &&
+      form.baseUrl.trim().length > 0 &&
+      (!isGoose || form.gooseBaseUrl.trim().length > 0),
+    [form.baseUrl, form.gooseBaseUrl, form.model, hasProvider, isGoose, isLoading, isSaving],
+  );
+  const canDiscover =
+    !isLoading &&
+    hasProvider &&
+    form.baseUrl.trim().length > 0 &&
+    (!requiresAPIKey || form.apiKey.length > 0 || (form.hasApiKey && !form.clearApiKey));
+  const modelOptions = useMemo(() => models.map((model) => ({ value: model.id, label: modelLabel(model) })), [models]);
+
+  const updateField = useCallback((field: keyof FormState, value: string | boolean | number) => {
+    setForm((current) => ({ ...current, [field]: value }));
+  }, []);
+
+  const handleProviderChange = useCallback((value: string) => {
+    const provider = Number(value) as LLMProvider;
+    discoveryRequestRef.current += 1;
+    setIsDiscovering(false);
+    setModels([]);
+    setModelError("");
+    setForm((current) => ({
+      ...current,
+      provider,
+      baseUrl: DEFAULT_BASE_URLS[provider] ?? "",
+      apiKey: "",
+      model: "",
+      hasApiKey: current.storedApiKeyProvider === provider,
+      clearApiKey: false,
+    }));
+  }, []);
+
+  const updateProviderConnection = useCallback((field: "apiKey" | "baseUrl", value: string) => {
+    discoveryRequestRef.current += 1;
+    setIsDiscovering(false);
+    setModels([]);
+    setModelError("");
+    setForm((current) => ({
+      ...current,
+      [field]: value,
+      model: "",
+      ...(field === "apiKey" && value ? { clearApiKey: false } : {}),
+    }));
+  }, []);
+
+  const toggleStoredAPIKey = useCallback(() => {
+    discoveryRequestRef.current += 1;
+    setIsDiscovering(false);
+    setModelError("");
+    setForm((current) => ({
+      ...current,
+      apiKey: "",
+      clearApiKey: !current.clearApiKey,
+    }));
+  }, []);
+
+  useEffect(() => {
+    if (!canDiscover) return;
+
+    const requestID = ++discoveryRequestRef.current;
+    const timeoutID = window.setTimeout(() => {
+      setIsDiscovering(true);
+      setModelError("");
+      void chatClient
+        .discoverModels({
+          provider: form.provider,
+          apiKey: form.apiKey,
+          baseUrl: form.baseUrl,
+          useStoredApiKey: requiresAPIKey && !form.apiKey && form.hasApiKey && !form.clearApiKey,
+        })
+        .then((response) => {
+          if (discoveryRequestRef.current !== requestID) return;
+          setModels(response.models);
+          setForm((current) => ({
+            ...current,
+            model: response.models.some((model) => model.id === current.model) ? current.model : "",
+          }));
+          if (response.models.length === 0) {
+            setModelError(
+              form.provider === LLMProvider.LLM_PROVIDER_OPENAI
+                ? "No OpenAI models compatible with the current agent flow are available for this API key."
+                : "The provider returned no available models.",
+            );
+          }
+        })
+        .catch((error: unknown) => {
+          if (discoveryRequestRef.current !== requestID) return;
+          setModelError(getErrorMessage(error, "Failed to fetch models from the provider."));
+        })
+        .finally(() => {
+          if (discoveryRequestRef.current === requestID) {
+            setIsDiscovering(false);
+          }
+        });
+    }, MODEL_DISCOVERY_DEBOUNCE_MS);
+
+    return () => {
+      window.clearTimeout(timeoutID);
+      if (discoveryRequestRef.current === requestID) {
+        discoveryRequestRef.current += 1;
+      }
+    };
+  }, [canDiscover, form.apiKey, form.baseUrl, form.clearApiKey, form.hasApiKey, form.provider, requiresAPIKey]);
+
+  const handleSave = useCallback(async () => {
+    if (!canSave) return;
+    setIsSaving(true);
+    try {
+      const response = await chatClient.updateLLMConfig({
+        harness: form.harness,
+        provider: form.provider,
+        apiKey: form.apiKey,
+        baseUrl: form.baseUrl,
+        model: form.model,
+        gooseBaseUrl: form.gooseBaseUrl,
+        gooseSecret: form.gooseSecret,
+        clearApiKey: form.clearApiKey,
+        clearGooseSecret: form.clearGooseSecret,
+      });
+      setForm(formFromConfig(response.config));
+      pushToast({ message: "Minerbot settings saved", status: STATUSES.success });
+    } catch (error) {
+      pushToast({
+        message: getErrorMessage(error, "Could not save Minerbot settings."),
+        status: STATUSES.error,
+      });
+    } finally {
+      setIsSaving(false);
+    }
+  }, [canSave, form]);
+
+  if (isLoading) {
+    return <div className="text-300 text-text-primary-50">Loading Minerbot settings...</div>;
+  }
+
+  return (
+    <section className="flex flex-col gap-5 rounded-xl border border-border-5 p-6" aria-labelledby="ai-settings-title">
+      <div>
+        <h2 id="ai-settings-title" className="text-heading-200 text-text-primary">
+          Minerbot
+        </h2>
+        <p className="mt-1 text-300 text-text-primary-50">
+          Choose the agent harness and bring your own model provider. Secrets are encrypted by the Fleet server and are
+          never returned to the browser.
+        </p>
+      </div>
+
+      {loadError ? (
+        <div role="alert" className="rounded-lg bg-intent-critical-10 p-3 text-300 text-text-critical">
+          {loadError}
+        </div>
+      ) : null}
+
+      <div className="grid grid-cols-2 gap-4 tablet:grid-cols-1">
+        <Select
+          id="ai-agent-harness"
+          label="Agent harness"
+          options={HARNESS_OPTIONS}
+          value={String(form.harness)}
+          onChange={(value) => updateField("harness", Number(value))}
+        />
+        <Select
+          id="ai-provider"
+          label="Model provider"
+          options={PROVIDER_OPTIONS}
+          value={String(form.provider)}
+          placeholder="Select a provider"
+          onChange={handleProviderChange}
+        />
+      </div>
+
+      <Input
+        id="ai-base-url"
+        label="Provider base URL"
+        initValue={form.baseUrl}
+        disabled={!hasProvider}
+        onChange={(value) => updateProviderConnection("baseUrl", value)}
+        required
+      />
+
+      {!hasProvider ? (
+        <p className="text-200 text-text-primary-50">Select a model provider to configure its connection.</p>
+      ) : requiresAPIKey ? (
+        <div className="flex items-center gap-3">
+          <div className="min-w-0 flex-1">
+            <Input
+              id="ai-api-key"
+              label={form.hasApiKey && !form.clearApiKey ? "API key (stored; enter to replace)" : "API key"}
+              type="password"
+              initValue={form.apiKey}
+              onChange={(value) => updateProviderConnection("apiKey", value)}
+            />
+          </div>
+          {form.hasApiKey ? (
+            <Button
+              variant={variants.textOnly}
+              size={sizes.compact}
+              text={form.clearApiKey ? "Keep stored key" : "Remove key"}
+              onClick={toggleStoredAPIKey}
+            />
+          ) : null}
+        </div>
+      ) : (
+        <p className="text-200 text-text-primary-50">Ollama does not require an API key by default.</p>
+      )}
+
+      <Select
+        id="ai-model"
+        label="Model"
+        options={modelOptions}
+        value={form.model}
+        placeholder={
+          !hasProvider
+            ? "Select a provider first"
+            : requiresAPIKey && !form.apiKey && (!form.hasApiKey || form.clearApiKey)
+              ? "Enter an API key to load models"
+              : isDiscovering
+                ? "Loading models..."
+                : models.length === 0
+                  ? "Waiting for models..."
+                  : "Select a model"
+        }
+        emptyMessage="No models available"
+        disabled={!hasProvider || isDiscovering || models.length === 0}
+        onChange={(value) => updateField("model", value)}
+      />
+
+      {modelError ? (
+        <div role="alert" className="rounded-lg bg-intent-critical-10 p-3 text-300 text-text-critical">
+          {modelError}
+        </div>
+      ) : null}
+
+      {isGoose ? (
+        <div className="flex flex-col gap-4 rounded-xl bg-intent-warning-10 p-4">
+          <p className="text-300 text-text-warning">
+            Goose ACP settings can be saved now, but chat remains unavailable until the server-side ACP adapter is
+            enabled.
+          </p>
+          <Input
+            id="goose-base-url"
+            label="Goose ACP base URL"
+            initValue={form.gooseBaseUrl}
+            onChange={(value) => updateField("gooseBaseUrl", value)}
+            required
+          />
+          <div className="flex items-center gap-3">
+            <div className="min-w-0 flex-1">
+              <Input
+                id="goose-secret"
+                label={
+                  form.hasGooseSecret && !form.clearGooseSecret
+                    ? "Goose secret (stored; enter to replace)"
+                    : "Goose secret"
+                }
+                type="password"
+                initValue={form.gooseSecret}
+                onChange={(value) => {
+                  updateField("gooseSecret", value);
+                  if (value) updateField("clearGooseSecret", false);
+                }}
+              />
+            </div>
+            {form.hasGooseSecret ? (
+              <Button
+                variant={variants.textOnly}
+                size={sizes.compact}
+                text={form.clearGooseSecret ? "Keep stored secret" : "Remove secret"}
+                onClick={() => updateField("clearGooseSecret", !form.clearGooseSecret)}
+              />
+            ) : null}
+          </div>
+        </div>
+      ) : null}
+
+      <div className="flex justify-end">
+        <Button
+          variant={variants.primary}
+          size={sizes.compact}
+          text="Save Minerbot settings"
+          disabled={!canSave}
+          loading={isSaving}
+          onClick={() => void handleSave()}
+        />
+      </div>
+    </section>
+  );
+};
+
+export default LLMProviderSettings;
diff --git a/client/src/protoFleet/features/settings/agents/index.ts b/client/src/protoFleet/features/settings/agents/index.ts
new file mode 100644
index 000000000..dbbd41e49
--- /dev/null
+++ b/client/src/protoFleet/features/settings/agents/index.ts
@@ -0,0 +1 @@
+export { default } from "./Agents";
diff --git a/client/src/protoFleet/routePrefetch.test.ts b/client/src/protoFleet/routePrefetch.test.ts
index 485bff116..576c60b0e 100644
--- a/client/src/protoFleet/routePrefetch.test.ts
+++ b/client/src/protoFleet/routePrefetch.test.ts
@@ -1,9 +1,25 @@
 import { describe, expect, it } from "vitest";
 
-import { importSettingsCurtailment, settingsRoutePrefetch } from "@/protoFleet/routePrefetch";
+import {
+  globalRoutePrefetch,
+  importMinerbotPage,
+  importSettingsAgents,
+  importSettingsCurtailment,
+  settingsRoutePrefetch,
+} from "@/protoFleet/routePrefetch";
+
+describe("globalRoutePrefetch", () => {
+  it("warms the Minerbot primary navigation page", () => {
+    expect(globalRoutePrefetch).toContain(importMinerbotPage);
+  });
+});
 
 describe("settingsRoutePrefetch", () => {
   it("warms the URL-only curtailment settings page", () => {
     expect(settingsRoutePrefetch).toContain(importSettingsCurtailment);
   });
+
+  it("warms the Agents settings page", () => {
+    expect(settingsRoutePrefetch).toContain(importSettingsAgents);
+  });
 });
diff --git a/client/src/protoFleet/routePrefetch.ts b/client/src/protoFleet/routePrefetch.ts
index 9c3a3b11e..40a3e2366 100644
--- a/client/src/protoFleet/routePrefetch.ts
+++ b/client/src/protoFleet/routePrefetch.ts
@@ -15,6 +15,7 @@ import type { RouteImporter } from "@/shared/utils/prefetchRoutes";
 export { singleMinerRoutePrefetch };
 
 export const importDashboard = () => import("@/protoFleet/features/dashboard/pages/Dashboard");
+export const importMinerbotPage = () => import("@/protoFleet/features/aiChat/pages/MinerbotPage");
 export const importMiners = () => import("./features/fleetManagement/components/Fleet");
 export const importActivityPage = () => import("@/protoFleet/features/activity/pages/ActivityPage");
 export const importEnergyPage = () => import("@/protoFleet/features/energy/pages/EnergyPage");
@@ -41,6 +42,7 @@ export const importSettingsSchedules = () =>
   import("@/protoFleet/features/settings/components/Schedules/SchedulesPage");
 export const importSettingsCurtailment = () => import("@/protoFleet/features/settings/components/Curtailment");
 export const importSettingsAlerts = () => import("@/protoFleet/features/alerts/pages/Alerts");
+export const importSettingsAgents = () => import("@/protoFleet/features/settings/agents");
 export const importSettingsIntegrations = () => import("@/protoFleet/features/settings/components/ApiKeys");
 export const importSiteDetailPage = () => import("@/protoFleet/features/sites/pages/SiteDetailPage");
 export const importBuildingPage = () => import("@/protoFleet/features/buildings/pages/BuildingPage");
@@ -63,6 +65,7 @@ export const globalRoutePrefetch: readonly RouteImporter[] = [
   importGroupsPage,
   importEnergyPage,
   importActivityPage,
+  importMinerbotPage,
   importSettingsLayout,
   importSettingsNetwork,
 ];
@@ -79,6 +82,7 @@ export const settingsRoutePrefetch: readonly RouteImporter[] = [
   importSettingsSchedules,
   importSettingsCurtailment,
   importSettingsAlerts,
+  importSettingsAgents,
   importSettingsIntegrations,
   importSettingsPreferences,
   importServerLogsPage,
diff --git a/client/src/protoFleet/router.tsx b/client/src/protoFleet/router.tsx
index 65e47dff9..043b43858 100644
--- a/client/src/protoFleet/router.tsx
+++ b/client/src/protoFleet/router.tsx
@@ -17,6 +17,7 @@ import {
   importFleetSitesPage,
   importGroupOverviewPage,
   importGroupsPage,
+  importMinerbotPage,
   importMiners,
   importMinersPage,
   importOnboardingSettingsPage,
@@ -24,6 +25,7 @@ import {
   importRacksPage,
   importSecurityPage,
   importServerLogsPage,
+  importSettingsAgents,
   importSettingsAlerts,
   importSettingsAuth,
   importSettingsCurtailment,
@@ -64,6 +66,7 @@ import { routerConfig as singleMinerRoutes } from "@/protoOS/router";
 // through this file. Auth metadata for the router lives in `routeAuth.ts`.
 
 const Dashboard = lazy(importDashboard);
+const MinerbotPage = lazy(importMinerbotPage);
 const Miners = lazy(importMiners);
 const ActivityPage = lazy(importActivityPage);
 const EnergyPage = lazy(importEnergyPage);
@@ -89,6 +92,7 @@ const SettingsNodes = lazy(importSettingsNodes);
 const SettingsSchedules = lazy(importSettingsSchedules);
 const SettingsCurtailment = lazy(importSettingsCurtailment);
 const SettingsAlerts = lazy(importSettingsAlerts);
+const SettingsAgents = lazy(importSettingsAgents);
 const SettingsIntegrations = lazy(importSettingsIntegrations);
 const SiteDetailPage = lazy(importSiteDetailPage);
 const BuildingPage = lazy(importBuildingPage);
@@ -246,6 +250,7 @@ const router = createBrowserRouter([
 
   createRoute("/racks/:rackId", <RackOverviewPage />, { hideShellHeader: true }),
   createRoute("/groups/:groupLabel", <GroupOverviewPage />, { hideShellHeader: true }),
+  createRoute("/minerbot", <MinerbotPage />),
 
   // /sites redirects into /fleet/sites.
   { path: "/sites", loader: sitesRedirectLoader },
@@ -344,6 +349,12 @@ const router = createBrowserRouter([
     path: "/settings/api-keys",
     loader: () => redirect("/settings/integrations"),
   },
+  createRoute(
+    "/settings/agents",
+    <SettingsLayout>
+      <SettingsAgents />
+    </SettingsLayout>,
+  ),
   createRoute(
     "/settings/integrations",
     <SettingsLayout>
diff --git a/client/src/shared/assets/icons/AI.tsx b/client/src/shared/assets/icons/AI.tsx
new file mode 100644
index 000000000..e013ac45b
--- /dev/null
+++ b/client/src/shared/assets/icons/AI.tsx
@@ -0,0 +1,64 @@
+import clsx from "clsx";
+
+import { IconProps } from "./types";
+
+type AIIconProps = IconProps & {
+  innerShadow?: boolean;
+};
+
+const path = (
+  <path
+    d="M7.40019 2.31649L7.91063 0.785172C8.2596 -0.261722 9.7404 -0.261725 10.0894 0.785169L10.5998 2.31649C11.3999 4.71669 13.2833 6.60013 15.6835 7.40019L17.2148 7.91063C18.2617 8.2596 18.2617 9.7404 17.2148 10.0894L15.6835 10.5998C13.2833 11.3999 11.3999 13.2833 10.5998 15.6835L10.0894 17.2148C9.7404 18.2617 8.2596 18.2617 7.91063 17.2148L7.40019 15.6835C6.60013 13.2833 4.71669 11.3999 2.31649 10.5998L0.785172 10.0894C-0.261722 9.7404 -0.261725 8.2596 0.785169 7.91063L2.31649 7.40019C4.71669 6.60013 6.60013 4.71669 7.40019 2.31649Z"
+    fill="currentColor"
+  />
+);
+
+const AI = ({ className, innerShadow = true, width = "w-[18px]", testId }: AIIconProps) => {
+  return (
+    <div className={clsx(width, className)} data-testid={testId}>
+      <svg
+        width="100%"
+        height="100%"
+        viewBox="0 0 18 18"
+        fill="none"
+        xmlns="http://www.w3.org/2000/svg"
+        preserveAspectRatio="xMidYMid meet"
+      >
+        {innerShadow ? (
+          <>
+            <g filter="url(#ai-icon-inner-shadow)">{path}</g>
+            <defs>
+              <filter
+                id="ai-icon-inner-shadow"
+                x="0"
+                y="0"
+                width="18"
+                height="21.2283"
+                filterUnits="userSpaceOnUse"
+                colorInterpolationFilters="sRGB"
+              >
+                <feFlood floodOpacity="0" result="BackgroundImageFix" />
+                <feBlend mode="normal" in="SourceGraphic" in2="BackgroundImageFix" result="shape" />
+                <feColorMatrix
+                  in="SourceAlpha"
+                  type="matrix"
+                  values="0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 127 0"
+                  result="hardAlpha"
+                />
+                <feOffset dy="3.22825" />
+                <feGaussianBlur stdDeviation="1.61413" />
+                <feComposite in2="hardAlpha" operator="arithmetic" k2="-1" k3="1" />
+                <feColorMatrix type="matrix" values="0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0.25 0" />
+                <feBlend mode="normal" in2="shape" result="effect1_innerShadow" />
+              </filter>
+            </defs>
+          </>
+        ) : (
+          path
+        )}
+      </svg>
+    </div>
+  );
+};
+
+export default AI;
diff --git a/client/src/shared/assets/icons/AIStroked.tsx b/client/src/shared/assets/icons/AIStroked.tsx
new file mode 100644
index 000000000..247572a5b
--- /dev/null
+++ b/client/src/shared/assets/icons/AIStroked.tsx
@@ -0,0 +1,28 @@
+import clsx from "clsx";
+
+import { iconSizes } from "./constants";
+import { IconProps } from "./types";
+
+const AIStroked = ({ className, width = iconSizes.medium, testId }: IconProps) => {
+  return (
+    <div className={clsx(width, className)} data-testid={testId}>
+      <svg
+        width="100%"
+        height="100%"
+        viewBox="0 0 18 18"
+        fill="none"
+        xmlns="http://www.w3.org/2000/svg"
+        preserveAspectRatio="xMidYMid meet"
+      >
+        <path
+          d="M7.57795 3.0591L8.03167 1.69793C8.34186 0.767358 9.65813 0.767356 9.96832 1.69793L10.422 3.0591C11.1332 5.19261 12.8074 6.86678 14.9409 7.57795L16.3021 8.03167C17.2326 8.34186 17.2326 9.65813 16.3021 9.96832L14.9409 10.422C12.8074 11.1332 11.1332 12.8074 10.422 14.9409L9.96833 16.3021C9.65813 17.2326 8.34187 17.2326 8.03167 16.3021L7.57795 14.9409C6.86678 12.8074 5.19261 11.1332 3.0591 10.422L1.69793 9.96833C0.767358 9.65813 0.767356 8.34187 1.69793 8.03167L3.0591 7.57795C5.19261 6.86678 6.86678 5.19261 7.57795 3.0591Z"
+          stroke="currentColor"
+          strokeWidth="2"
+          strokeLinejoin="round"
+        />
+      </svg>
+    </div>
+  );
+};
+
+export default AIStroked;
diff --git a/client/src/shared/assets/icons/index.tsx b/client/src/shared/assets/icons/index.tsx
index 0929a3c4a..f50839917 100644
--- a/client/src/shared/assets/icons/index.tsx
+++ b/client/src/shared/assets/icons/index.tsx
@@ -1,4 +1,6 @@
 import Activity from "./Activity";
+import AI from "./AI";
+import AIStroked from "./AIStroked";
 import Alert from "./Alert";
 import ArrowDown from "./ArrowDown";
 import ArrowLeftCompact from "./ArrowLeftCompact";
@@ -88,6 +90,8 @@ import Unpair from "./Unpair";
 
 export {
   Activity,
+  AI,
+  AIStroked,
   Alert,
   ArrowDown,
   ArrowLeftCompact,
diff --git a/deployment-files/docker-compose.yaml b/deployment-files/docker-compose.yaml
index 0fec26520..77c942744 100644
--- a/deployment-files/docker-compose.yaml
+++ b/deployment-files/docker-compose.yaml
@@ -15,6 +15,7 @@ services:
       DB_PASSWORD: "${DB_PASSWORD}"
       DB_DSN: "${DB_DSN:-}"
       ENCRYPT_SERVICE_MASTER_KEY: "${ENCRYPT_SERVICE_MASTER_KEY}"
+      CHAT_ALLOW_PRIVATE_OLLAMA: "${CHAT_ALLOW_PRIVATE_OLLAMA:-false}"
       SESSION_COOKIE_SECURE: "${SESSION_COOKIE_SECURE:-true}"
       PLUGINS_DIR: "/app/plugins"
       PLUGINS_ENABLED: "true"
diff --git a/proto/chat/v1/chat.proto b/proto/chat/v1/chat.proto
new file mode 100644
index 000000000..59f90cd88
--- /dev/null
+++ b/proto/chat/v1/chat.proto
@@ -0,0 +1,200 @@
+syntax = "proto3";
+
+package chat.v1;
+
+import "buf/validate/validate.proto";
+
+option go_package = "github.com/block/proto-fleet/server/generated/grpc/chat/v1;chatv1";
+
+// ChatService owns the organization-scoped BYOLLM configuration and the
+// streaming agent conversation boundary used by the Proto Fleet web client.
+service ChatService {
+  rpc GetLLMConfig(GetLLMConfigRequest) returns (GetLLMConfigResponse);
+  rpc DiscoverModels(DiscoverModelsRequest) returns (DiscoverModelsResponse);
+  rpc UpdateLLMConfig(UpdateLLMConfigRequest) returns (UpdateLLMConfigResponse);
+  rpc SendMessage(SendMessageRequest) returns (stream SendMessageResponse);
+  // ResolveToolConfirmation approves or cancels a pending write requested by
+  // SendMessage. The original stream remains open and resumes after this RPC.
+  rpc ResolveToolConfirmation(ResolveToolConfirmationRequest) returns (ResolveToolConfirmationResponse);
+}
+
+enum AgentHarness {
+  AGENT_HARNESS_UNSPECIFIED = 0;
+  AGENT_HARNESS_NATIVE = 1;
+  AGENT_HARNESS_GOOSE = 2;
+}
+
+enum LLMProvider {
+  LLM_PROVIDER_UNSPECIFIED = 0;
+  LLM_PROVIDER_OPENAI = 1;
+  LLM_PROVIDER_ANTHROPIC = 2;
+  LLM_PROVIDER_OLLAMA = 3;
+  LLM_PROVIDER_CUSTOM = 4;
+}
+
+message LLMConfig {
+  reserved 6;
+
+  AgentHarness harness = 1;
+  LLMProvider provider = 2;
+  bool has_api_key = 3;
+  string base_url = 4;
+  string model = 5;
+  string goose_base_url = 7;
+  bool has_goose_secret = 8;
+  bool configured = 9;
+}
+
+message GetLLMConfigRequest {}
+
+message GetLLMConfigResponse {
+  LLMConfig config = 1;
+}
+
+message DiscoverModelsRequest {
+  LLMProvider provider = 1 [(buf.validate.field).enum = {
+    defined_only: true
+    not_in: [0]
+  }];
+  // Write-only credential used for this discovery request. It is not saved.
+  string api_key = 2 [(buf.validate.field).string.max_len = 8192];
+  string base_url = 3 [(buf.validate.field).string.max_len = 2048];
+  // Uses the encrypted credential already saved for this provider. This cannot
+  // be combined with api_key.
+  bool use_stored_api_key = 4;
+}
+
+message AvailableModel {
+  string id = 1;
+  string display_name = 2;
+}
+
+message DiscoverModelsResponse {
+  repeated AvailableModel models = 1;
+}
+
+message UpdateLLMConfigRequest {
+  reserved 6;
+
+  AgentHarness harness = 1 [(buf.validate.field).enum = {
+    defined_only: true
+    not_in: [0]
+  }];
+  LLMProvider provider = 2 [(buf.validate.field).enum = {
+    defined_only: true
+    not_in: [0]
+  }];
+  // Write-only. An empty value preserves the existing key unless
+  // clear_api_key is true.
+  string api_key = 3 [(buf.validate.field).string.max_len = 8192];
+  string base_url = 4 [(buf.validate.field).string.max_len = 2048];
+  string model = 5 [(buf.validate.field).string = {
+    min_len: 1
+    max_len: 255
+  }];
+  string goose_base_url = 7 [(buf.validate.field).string.max_len = 2048];
+  // Write-only shared secret for a remote Goose ACP endpoint.
+  string goose_secret = 8 [(buf.validate.field).string.max_len = 8192];
+  bool clear_api_key = 9;
+  bool clear_goose_secret = 10;
+}
+
+message UpdateLLMConfigResponse {
+  LLMConfig config = 1;
+}
+
+enum ChatRole {
+  CHAT_ROLE_UNSPECIFIED = 0;
+  CHAT_ROLE_USER = 1;
+  CHAT_ROLE_ASSISTANT = 2;
+}
+
+message ChatTurn {
+  ChatRole role = 1 [(buf.validate.field).enum = {
+    defined_only: true
+    not_in: [0]
+  }];
+  string content = 2 [(buf.validate.field).string = {
+    min_len: 1
+    max_len: 32768
+  }];
+}
+
+message SendMessageRequest {
+  string conversation_id = 1 [(buf.validate.field).string.max_len = 128];
+  string content = 2 [(buf.validate.field).string = {
+    min_len: 1
+    max_len: 32768
+  }];
+  // Conversations remain client-side. Supplying prior turns keeps the server
+  // stateless while preserving multi-turn context.
+  repeated ChatTurn history = 3 [(buf.validate.field).repeated.max_items = 50];
+}
+
+message TextDelta {
+  string content = 1;
+}
+
+message ToolCall {
+  string id = 1;
+  string name = 2;
+  string summary = 3;
+}
+
+message ToolResult {
+  string id = 1;
+  string name = 2;
+  bool success = 3;
+  string summary = 4;
+  bool cancelled = 5;
+}
+
+message ToolConfirmationDetail {
+  string label = 1;
+  string value = 2;
+}
+
+message ToolConfirmationRequired {
+  // Opaque, short-lived identifier bound to the current operator and org.
+  string confirmation_id = 1;
+  string tool_call_id = 2;
+  string tool_name = 3;
+  string title = 4;
+  string description = 5;
+  string confirm_label = 6;
+  repeated ToolConfirmationDetail details = 7;
+}
+
+enum ToolConfirmationDecision {
+  TOOL_CONFIRMATION_DECISION_UNSPECIFIED = 0;
+  TOOL_CONFIRMATION_DECISION_APPROVE = 1;
+  TOOL_CONFIRMATION_DECISION_CANCEL = 2;
+}
+
+message ResolveToolConfirmationRequest {
+  string confirmation_id = 1 [(buf.validate.field).string = {
+    min_len: 1
+    max_len: 128
+  }];
+  ToolConfirmationDecision decision = 2 [(buf.validate.field).enum = {
+    defined_only: true
+    not_in: [0]
+  }];
+}
+
+message ResolveToolConfirmationResponse {}
+
+message Done {
+  string finish_reason = 1;
+}
+
+message SendMessageResponse {
+  string message_id = 1;
+  oneof event {
+    TextDelta text_delta = 2;
+    ToolCall tool_call = 3;
+    ToolResult tool_result = 4;
+    Done done = 5;
+    ToolConfirmationRequired confirmation_required = 6;
+  }
+}
diff --git a/server/cmd/fleetd/config.go b/server/cmd/fleetd/config.go
index 69ef941e0..2035a6a6c 100644
--- a/server/cmd/fleetd/config.go
+++ b/server/cmd/fleetd/config.go
@@ -3,6 +3,7 @@ package main
 import (
 	"time"
 
+	"github.com/block/proto-fleet/server/internal/domain/chat"
 	"github.com/block/proto-fleet/server/internal/domain/command"
 	curtailmentReconciler "github.com/block/proto-fleet/server/internal/domain/curtailment/reconciler"
 	"github.com/block/proto-fleet/server/internal/domain/diagnostics"
@@ -42,6 +43,7 @@ type Config struct {
 	Log            logging.Config               `embed:"" prefix:"logging-" envprefix:"LOG_"`
 	HTTP           HTTPConfig                   `embed:"" prefix:"http-" envprefix:"HTTP_"`
 	Auth           token.Config                 `embed:"" prefix:"auth-" envprefix:"AUTH_"`
+	Chat           chat.ProviderEgressConfig    `embed:"" prefix:"chat-" envprefix:"CHAT_"`
 	Session        session.Config               `embed:"" prefix:"session-" envprefix:"SESSION_"`
 	Pools          pools.Config                 `embed:"" prefix:"pools-" envprefix:"POOLS_"`
 	Encrypt        encrypt.Config               `embed:"" prefix:"encrypt-" envprefix:"ENCRYPT_"`
diff --git a/server/cmd/fleetd/main.go b/server/cmd/fleetd/main.go
index e994fd6ee..7ef804813 100644
--- a/server/cmd/fleetd/main.go
+++ b/server/cmd/fleetd/main.go
@@ -45,6 +45,7 @@ import (
 	"github.com/block/proto-fleet/server/generated/grpc/auth/v1/authv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/authz/v1/authzv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/buildings/v1/buildingsv1connect"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/collection/v1/collectionv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/curtailment/v1/curtailmentv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/device_set/v1/device_setv1connect"
@@ -69,6 +70,7 @@ import (
 	apikeyDomain "github.com/block/proto-fleet/server/internal/domain/apikey"
 	authDomain "github.com/block/proto-fleet/server/internal/domain/auth"
 	buildingsDomain "github.com/block/proto-fleet/server/internal/domain/buildings"
+	chatDomain "github.com/block/proto-fleet/server/internal/domain/chat"
 	collectionDomain "github.com/block/proto-fleet/server/internal/domain/collection"
 	commandDomain "github.com/block/proto-fleet/server/internal/domain/command"
 	curtailmentDomain "github.com/block/proto-fleet/server/internal/domain/curtailment"
@@ -102,6 +104,7 @@ import (
 	"github.com/block/proto-fleet/server/internal/handlers/auth"
 	authzHandler "github.com/block/proto-fleet/server/internal/handlers/authz"
 	buildingsHandler "github.com/block/proto-fleet/server/internal/handlers/buildings"
+	chatHandler "github.com/block/proto-fleet/server/internal/handlers/chat"
 	collectionHandler "github.com/block/proto-fleet/server/internal/handlers/collection"
 	"github.com/block/proto-fleet/server/internal/handlers/command"
 	curtailmentHandler "github.com/block/proto-fleet/server/internal/handlers/curtailment"
@@ -171,6 +174,7 @@ var reflectEnabledServices = []string{
 	sitemapv1connect.SiteMapServiceName,
 	curtailmentv1connect.CurtailmentServiceName,
 	device_setv1connect.DeviceSetServiceName,
+	chatv1connect.ChatServiceName,
 }
 
 func start(config *Config) error {
@@ -649,6 +653,11 @@ func start(config *Config) error {
 	alertChannelStore := sqlstores.NewSQLAlertChannelStore(conn)
 	alertsDeliverer := alertsDomain.NewDeliverer(alertChannelStore, encryptSvc, alertChannelStore, config.Metrics.AlertDestinations, config.PublicURL)
 	alertsSvc := alertsDomain.NewService(grafanaClient, alertChannelStore, encryptSvc, alertsDeliverer, config.Metrics.AlertDestinations)
+	llmConfigStore := sqlstores.NewSQLLLMConfigStore(conn)
+	llmConfigSvc := chatDomain.NewConfigService(llmConfigStore, encryptSvc, config.Chat)
+	chatModelClient := chatDomain.NewHTTPModelClient(config.Chat)
+	chatConfirmationBroker := chatDomain.NewConfirmationBroker()
+	chatAgent := chatDomain.NewAgent(chatModelClient, chatConfirmationBroker)
 
 	middlewares := []server.Middleware{
 		middleware.NewCORSMiddleware(config.HTTP.SuppressCors),
@@ -705,12 +714,19 @@ func start(config *Config) error {
 	mux.Handle(onboardingv1connect.NewOnboardingServiceHandler(onboarding.NewHandler(authSvc, onboardingSvc), li))
 	mux.Handle(pairingv1connect.NewPairingServiceHandler(pairing.NewHandler(pairingSvc, fleetNodeDiscoverySvc, fleetNodePairingSvc), li))
 	mux.Handle(networkinfov1connect.NewNetworkInfoServiceHandler(networkinfo.NewHandler(pairingSvc), li))
-	mux.Handle(fleetmanagementv1connect.NewFleetManagementServiceHandler(fleetmanagement.NewHandler(fleetMgmtSvc), li))
-	mux.Handle(minercommandv1connect.NewMinerCommandServiceHandler(command.NewHandler(commandSvc), li))
-	mux.Handle(poolsv1connect.NewPoolsServiceHandler(pools.NewHandler(poolsSvc), li))
-	mux.Handle(schedulev1connect.NewScheduleServiceHandler(scheduleHandler.NewHandler(scheduleSvc), li))
+	fleetManagementHandler := fleetmanagement.NewHandler(fleetMgmtSvc)
+	poolsHandler := pools.NewHandler(poolsSvc)
+	siteServiceHandler := sitesHandler.NewHandler(sitesSvc)
+	deviceSetServiceHandler := devicesetHandler.NewHandler(collectionSvc)
+	commandServiceHandler := command.NewHandler(commandSvc)
+	scheduleServiceHandler := scheduleHandler.NewHandler(scheduleSvc)
+
+	mux.Handle(fleetmanagementv1connect.NewFleetManagementServiceHandler(fleetManagementHandler, li))
+	mux.Handle(minercommandv1connect.NewMinerCommandServiceHandler(commandServiceHandler, li))
+	mux.Handle(poolsv1connect.NewPoolsServiceHandler(poolsHandler, li))
+	mux.Handle(schedulev1connect.NewScheduleServiceHandler(scheduleServiceHandler, li))
 	mux.Handle(curtailmentv1connect.NewCurtailmentServiceHandler(curtailmentHandler.NewHandlerWithAutomation(curtailmentSvc, curtailmentResponseProfileSvc, curtailmentAutomationSvc, mqttSettingsSvc), li))
-	mux.Handle(sitesv1connect.NewSiteServiceHandler(sitesHandler.NewHandler(sitesSvc), li))
+	mux.Handle(sitesv1connect.NewSiteServiceHandler(siteServiceHandler, li))
 	mux.Handle(buildingsv1connect.NewBuildingServiceHandler(buildingsHandler.NewHandler(buildingsSvc), li))
 	mux.Handle(infrastructurev1connect.NewInfrastructureServiceHandler(infrastructureHandler.NewHandler(infrastructureSvc), li))
 	mux.Handle(sitemapv1connect.NewSiteMapServiceHandler(
@@ -725,7 +741,7 @@ func start(config *Config) error {
 	))
 	mux.Handle(fleetnodeadminv1connect.NewFleetNodeAdminServiceHandler(admin.NewHandler(fleetNodeEnrollmentSvc, fleetNodePairingSvc, fleetNodeDiscoverySvc), li))
 	mux.Handle(collectionv1connect.NewDeviceCollectionServiceHandler(collectionHandler.NewHandler(collectionSvc), li))
-	mux.Handle(device_setv1connect.NewDeviceSetServiceHandler(devicesetHandler.NewHandler(collectionSvc), li))
+	mux.Handle(device_setv1connect.NewDeviceSetServiceHandler(deviceSetServiceHandler, li))
 	mux.Handle(telemetryv1connect.NewTelemetryServiceHandler(telemetryHandler.NewHandler(telemetryService), li))
 	mux.Handle(errorsv1connect.NewErrorQueryServiceHandler(errorqueryHandler.NewHandler(diagnosticsService), li))
 	mux.Handle(foremanimportv1connect.NewForemanImportServiceHandler(foremanImportHandler.NewHandler(foremanImportSvc), li))
@@ -733,6 +749,13 @@ func start(config *Config) error {
 	mux.Handle(apikeyv1connect.NewApiKeyServiceHandler(apikeyHandler.NewHandler(apiKeySvc), li))
 	mux.Handle(authzv1connect.NewAuthzServiceHandler(authzHandler.NewHandler(authz.NewService(conn, activitySvc)), li))
 	mux.Handle(serverlogv1connect.NewServerLogServiceHandler(serverlogHandler.NewHandler(logging.DefaultBuffer()), li))
+	mux.Handle(chatv1connect.NewChatServiceHandler(chatHandler.NewHandler(
+		llmConfigSvc,
+		chatAgent,
+		chatModelClient,
+		chatHandler.NewFleetTools(fleetManagementHandler, siteServiceHandler, poolsHandler, deviceSetServiceHandler, commandServiceHandler, scheduleServiceHandler),
+		chatConfirmationBroker,
+	), li))
 
 	alertHandler := alertsHandler.NewHandler(alertsSvc, notificationHistoryStore)
 	mux.Handle(alertsv1connect.NewChannelServiceHandler(alertHandler, li))
diff --git a/server/cmd/fleetnode/telemetry.go b/server/cmd/fleetnode/telemetry.go
index 4fbcd0585..7c2948048 100644
--- a/server/cmd/fleetnode/telemetry.go
+++ b/server/cmd/fleetnode/telemetry.go
@@ -260,15 +260,16 @@ type telemetryDeviceCloser interface {
 }
 
 func closeTelemetryDeviceAsync(device telemetryDeviceCloser) bool {
+	closeTokens := telemetryDeviceCloseTokens
 	select {
-	case telemetryDeviceCloseTokens <- struct{}{}:
+	case closeTokens <- struct{}{}:
 	default:
 		slog.Warn("telemetry device close worker capacity is exhausted; closing synchronously")
 		closeTelemetryDevice(device)
 		return false
 	}
 	go func() {
-		defer func() { <-telemetryDeviceCloseTokens }()
+		defer func() { <-closeTokens }()
 		closeTelemetryDevice(device)
 	}()
 	return true
diff --git a/server/docker-compose.yaml b/server/docker-compose.yaml
index 6564170f0..9e3948687 100644
--- a/server/docker-compose.yaml
+++ b/server/docker-compose.yaml
@@ -40,6 +40,7 @@ services:
       DB_PASSWORD: "fleet"
       ENCRYPT_SERVICE_MASTER_KEY: "EV0g7BoFQnqshvzep9knZsUUmvsqSBsjDAJus7ri0B8="
       INFRASTRUCTURE_OT_CONTROL_SUBNETS: "${INFRASTRUCTURE_OT_CONTROL_SUBNETS:-}"
+      CHAT_ALLOW_PRIVATE_OLLAMA: "${CHAT_ALLOW_PRIVATE_OLLAMA:-false}"
       PLUGINS_DIR: "/app/plugins"
       PLUGINS_ENABLED: "true"
       PROTO_MINER_PORT: "8080"
diff --git a/server/generated/grpc/chat/v1/chat.pb.go b/server/generated/grpc/chat/v1/chat.pb.go
new file mode 100644
index 000000000..9f3f7adde
--- /dev/null
+++ b/server/generated/grpc/chat/v1/chat.pb.go
@@ -0,0 +1,1791 @@
+// Code generated by protoc-gen-go. DO NOT EDIT.
+// versions:
+// 	protoc-gen-go v1.36.5
+// 	protoc        (unknown)
+// source: chat/v1/chat.proto
+
+package chatv1
+
+import (
+	reflect "reflect"
+	sync "sync"
+	unsafe "unsafe"
+
+	_ "buf.build/gen/go/bufbuild/protovalidate/protocolbuffers/go/buf/validate"
+	protoreflect "google.golang.org/protobuf/reflect/protoreflect"
+	protoimpl "google.golang.org/protobuf/runtime/protoimpl"
+)
+
+const (
+	// Verify that this generated code is sufficiently up-to-date.
+	_ = protoimpl.EnforceVersion(20 - protoimpl.MinVersion)
+	// Verify that runtime/protoimpl is sufficiently up-to-date.
+	_ = protoimpl.EnforceVersion(protoimpl.MaxVersion - 20)
+)
+
+type AgentHarness int32
+
+const (
+	AgentHarness_AGENT_HARNESS_UNSPECIFIED AgentHarness = 0
+	AgentHarness_AGENT_HARNESS_NATIVE      AgentHarness = 1
+	AgentHarness_AGENT_HARNESS_GOOSE       AgentHarness = 2
+)
+
+// Enum value maps for AgentHarness.
+var (
+	AgentHarness_name = map[int32]string{
+		0: "AGENT_HARNESS_UNSPECIFIED",
+		1: "AGENT_HARNESS_NATIVE",
+		2: "AGENT_HARNESS_GOOSE",
+	}
+	AgentHarness_value = map[string]int32{
+		"AGENT_HARNESS_UNSPECIFIED": 0,
+		"AGENT_HARNESS_NATIVE":      1,
+		"AGENT_HARNESS_GOOSE":       2,
+	}
+)
+
+func (x AgentHarness) Enum() *AgentHarness {
+	p := new(AgentHarness)
+	*p = x
+	return p
+}
+
+func (x AgentHarness) String() string {
+	return protoimpl.X.EnumStringOf(x.Descriptor(), protoreflect.EnumNumber(x))
+}
+
+func (AgentHarness) Descriptor() protoreflect.EnumDescriptor {
+	return file_chat_v1_chat_proto_enumTypes[0].Descriptor()
+}
+
+func (AgentHarness) Type() protoreflect.EnumType {
+	return &file_chat_v1_chat_proto_enumTypes[0]
+}
+
+func (x AgentHarness) Number() protoreflect.EnumNumber {
+	return protoreflect.EnumNumber(x)
+}
+
+// Deprecated: Use AgentHarness.Descriptor instead.
+func (AgentHarness) EnumDescriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{0}
+}
+
+type LLMProvider int32
+
+const (
+	LLMProvider_LLM_PROVIDER_UNSPECIFIED LLMProvider = 0
+	LLMProvider_LLM_PROVIDER_OPENAI      LLMProvider = 1
+	LLMProvider_LLM_PROVIDER_ANTHROPIC   LLMProvider = 2
+	LLMProvider_LLM_PROVIDER_OLLAMA      LLMProvider = 3
+	LLMProvider_LLM_PROVIDER_CUSTOM      LLMProvider = 4
+)
+
+// Enum value maps for LLMProvider.
+var (
+	LLMProvider_name = map[int32]string{
+		0: "LLM_PROVIDER_UNSPECIFIED",
+		1: "LLM_PROVIDER_OPENAI",
+		2: "LLM_PROVIDER_ANTHROPIC",
+		3: "LLM_PROVIDER_OLLAMA",
+		4: "LLM_PROVIDER_CUSTOM",
+	}
+	LLMProvider_value = map[string]int32{
+		"LLM_PROVIDER_UNSPECIFIED": 0,
+		"LLM_PROVIDER_OPENAI":      1,
+		"LLM_PROVIDER_ANTHROPIC":   2,
+		"LLM_PROVIDER_OLLAMA":      3,
+		"LLM_PROVIDER_CUSTOM":      4,
+	}
+)
+
+func (x LLMProvider) Enum() *LLMProvider {
+	p := new(LLMProvider)
+	*p = x
+	return p
+}
+
+func (x LLMProvider) String() string {
+	return protoimpl.X.EnumStringOf(x.Descriptor(), protoreflect.EnumNumber(x))
+}
+
+func (LLMProvider) Descriptor() protoreflect.EnumDescriptor {
+	return file_chat_v1_chat_proto_enumTypes[1].Descriptor()
+}
+
+func (LLMProvider) Type() protoreflect.EnumType {
+	return &file_chat_v1_chat_proto_enumTypes[1]
+}
+
+func (x LLMProvider) Number() protoreflect.EnumNumber {
+	return protoreflect.EnumNumber(x)
+}
+
+// Deprecated: Use LLMProvider.Descriptor instead.
+func (LLMProvider) EnumDescriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{1}
+}
+
+type ChatRole int32
+
+const (
+	ChatRole_CHAT_ROLE_UNSPECIFIED ChatRole = 0
+	ChatRole_CHAT_ROLE_USER        ChatRole = 1
+	ChatRole_CHAT_ROLE_ASSISTANT   ChatRole = 2
+)
+
+// Enum value maps for ChatRole.
+var (
+	ChatRole_name = map[int32]string{
+		0: "CHAT_ROLE_UNSPECIFIED",
+		1: "CHAT_ROLE_USER",
+		2: "CHAT_ROLE_ASSISTANT",
+	}
+	ChatRole_value = map[string]int32{
+		"CHAT_ROLE_UNSPECIFIED": 0,
+		"CHAT_ROLE_USER":        1,
+		"CHAT_ROLE_ASSISTANT":   2,
+	}
+)
+
+func (x ChatRole) Enum() *ChatRole {
+	p := new(ChatRole)
+	*p = x
+	return p
+}
+
+func (x ChatRole) String() string {
+	return protoimpl.X.EnumStringOf(x.Descriptor(), protoreflect.EnumNumber(x))
+}
+
+func (ChatRole) Descriptor() protoreflect.EnumDescriptor {
+	return file_chat_v1_chat_proto_enumTypes[2].Descriptor()
+}
+
+func (ChatRole) Type() protoreflect.EnumType {
+	return &file_chat_v1_chat_proto_enumTypes[2]
+}
+
+func (x ChatRole) Number() protoreflect.EnumNumber {
+	return protoreflect.EnumNumber(x)
+}
+
+// Deprecated: Use ChatRole.Descriptor instead.
+func (ChatRole) EnumDescriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{2}
+}
+
+type ToolConfirmationDecision int32
+
+const (
+	ToolConfirmationDecision_TOOL_CONFIRMATION_DECISION_UNSPECIFIED ToolConfirmationDecision = 0
+	ToolConfirmationDecision_TOOL_CONFIRMATION_DECISION_APPROVE     ToolConfirmationDecision = 1
+	ToolConfirmationDecision_TOOL_CONFIRMATION_DECISION_CANCEL      ToolConfirmationDecision = 2
+)
+
+// Enum value maps for ToolConfirmationDecision.
+var (
+	ToolConfirmationDecision_name = map[int32]string{
+		0: "TOOL_CONFIRMATION_DECISION_UNSPECIFIED",
+		1: "TOOL_CONFIRMATION_DECISION_APPROVE",
+		2: "TOOL_CONFIRMATION_DECISION_CANCEL",
+	}
+	ToolConfirmationDecision_value = map[string]int32{
+		"TOOL_CONFIRMATION_DECISION_UNSPECIFIED": 0,
+		"TOOL_CONFIRMATION_DECISION_APPROVE":     1,
+		"TOOL_CONFIRMATION_DECISION_CANCEL":      2,
+	}
+)
+
+func (x ToolConfirmationDecision) Enum() *ToolConfirmationDecision {
+	p := new(ToolConfirmationDecision)
+	*p = x
+	return p
+}
+
+func (x ToolConfirmationDecision) String() string {
+	return protoimpl.X.EnumStringOf(x.Descriptor(), protoreflect.EnumNumber(x))
+}
+
+func (ToolConfirmationDecision) Descriptor() protoreflect.EnumDescriptor {
+	return file_chat_v1_chat_proto_enumTypes[3].Descriptor()
+}
+
+func (ToolConfirmationDecision) Type() protoreflect.EnumType {
+	return &file_chat_v1_chat_proto_enumTypes[3]
+}
+
+func (x ToolConfirmationDecision) Number() protoreflect.EnumNumber {
+	return protoreflect.EnumNumber(x)
+}
+
+// Deprecated: Use ToolConfirmationDecision.Descriptor instead.
+func (ToolConfirmationDecision) EnumDescriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{3}
+}
+
+type LLMConfig struct {
+	state          protoimpl.MessageState `protogen:"open.v1"`
+	Harness        AgentHarness           `protobuf:"varint,1,opt,name=harness,proto3,enum=chat.v1.AgentHarness" json:"harness,omitempty"`
+	Provider       LLMProvider            `protobuf:"varint,2,opt,name=provider,proto3,enum=chat.v1.LLMProvider" json:"provider,omitempty"`
+	HasApiKey      bool                   `protobuf:"varint,3,opt,name=has_api_key,json=hasApiKey,proto3" json:"has_api_key,omitempty"`
+	BaseUrl        string                 `protobuf:"bytes,4,opt,name=base_url,json=baseUrl,proto3" json:"base_url,omitempty"`
+	Model          string                 `protobuf:"bytes,5,opt,name=model,proto3" json:"model,omitempty"`
+	GooseBaseUrl   string                 `protobuf:"bytes,7,opt,name=goose_base_url,json=gooseBaseUrl,proto3" json:"goose_base_url,omitempty"`
+	HasGooseSecret bool                   `protobuf:"varint,8,opt,name=has_goose_secret,json=hasGooseSecret,proto3" json:"has_goose_secret,omitempty"`
+	Configured     bool                   `protobuf:"varint,9,opt,name=configured,proto3" json:"configured,omitempty"`
+	unknownFields  protoimpl.UnknownFields
+	sizeCache      protoimpl.SizeCache
+}
+
+func (x *LLMConfig) Reset() {
+	*x = LLMConfig{}
+	mi := &file_chat_v1_chat_proto_msgTypes[0]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *LLMConfig) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*LLMConfig) ProtoMessage() {}
+
+func (x *LLMConfig) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[0]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use LLMConfig.ProtoReflect.Descriptor instead.
+func (*LLMConfig) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{0}
+}
+
+func (x *LLMConfig) GetHarness() AgentHarness {
+	if x != nil {
+		return x.Harness
+	}
+	return AgentHarness_AGENT_HARNESS_UNSPECIFIED
+}
+
+func (x *LLMConfig) GetProvider() LLMProvider {
+	if x != nil {
+		return x.Provider
+	}
+	return LLMProvider_LLM_PROVIDER_UNSPECIFIED
+}
+
+func (x *LLMConfig) GetHasApiKey() bool {
+	if x != nil {
+		return x.HasApiKey
+	}
+	return false
+}
+
+func (x *LLMConfig) GetBaseUrl() string {
+	if x != nil {
+		return x.BaseUrl
+	}
+	return ""
+}
+
+func (x *LLMConfig) GetModel() string {
+	if x != nil {
+		return x.Model
+	}
+	return ""
+}
+
+func (x *LLMConfig) GetGooseBaseUrl() string {
+	if x != nil {
+		return x.GooseBaseUrl
+	}
+	return ""
+}
+
+func (x *LLMConfig) GetHasGooseSecret() bool {
+	if x != nil {
+		return x.HasGooseSecret
+	}
+	return false
+}
+
+func (x *LLMConfig) GetConfigured() bool {
+	if x != nil {
+		return x.Configured
+	}
+	return false
+}
+
+type GetLLMConfigRequest struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *GetLLMConfigRequest) Reset() {
+	*x = GetLLMConfigRequest{}
+	mi := &file_chat_v1_chat_proto_msgTypes[1]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *GetLLMConfigRequest) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*GetLLMConfigRequest) ProtoMessage() {}
+
+func (x *GetLLMConfigRequest) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[1]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use GetLLMConfigRequest.ProtoReflect.Descriptor instead.
+func (*GetLLMConfigRequest) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{1}
+}
+
+type GetLLMConfigResponse struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Config        *LLMConfig             `protobuf:"bytes,1,opt,name=config,proto3" json:"config,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *GetLLMConfigResponse) Reset() {
+	*x = GetLLMConfigResponse{}
+	mi := &file_chat_v1_chat_proto_msgTypes[2]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *GetLLMConfigResponse) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*GetLLMConfigResponse) ProtoMessage() {}
+
+func (x *GetLLMConfigResponse) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[2]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use GetLLMConfigResponse.ProtoReflect.Descriptor instead.
+func (*GetLLMConfigResponse) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{2}
+}
+
+func (x *GetLLMConfigResponse) GetConfig() *LLMConfig {
+	if x != nil {
+		return x.Config
+	}
+	return nil
+}
+
+type DiscoverModelsRequest struct {
+	state    protoimpl.MessageState `protogen:"open.v1"`
+	Provider LLMProvider            `protobuf:"varint,1,opt,name=provider,proto3,enum=chat.v1.LLMProvider" json:"provider,omitempty"`
+	// Write-only credential used for this discovery request. It is not saved.
+	ApiKey  string `protobuf:"bytes,2,opt,name=api_key,json=apiKey,proto3" json:"api_key,omitempty"`
+	BaseUrl string `protobuf:"bytes,3,opt,name=base_url,json=baseUrl,proto3" json:"base_url,omitempty"`
+	// Uses the encrypted credential already saved for this provider. This cannot
+	// be combined with api_key.
+	UseStoredApiKey bool `protobuf:"varint,4,opt,name=use_stored_api_key,json=useStoredApiKey,proto3" json:"use_stored_api_key,omitempty"`
+	unknownFields   protoimpl.UnknownFields
+	sizeCache       protoimpl.SizeCache
+}
+
+func (x *DiscoverModelsRequest) Reset() {
+	*x = DiscoverModelsRequest{}
+	mi := &file_chat_v1_chat_proto_msgTypes[3]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *DiscoverModelsRequest) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*DiscoverModelsRequest) ProtoMessage() {}
+
+func (x *DiscoverModelsRequest) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[3]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use DiscoverModelsRequest.ProtoReflect.Descriptor instead.
+func (*DiscoverModelsRequest) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{3}
+}
+
+func (x *DiscoverModelsRequest) GetProvider() LLMProvider {
+	if x != nil {
+		return x.Provider
+	}
+	return LLMProvider_LLM_PROVIDER_UNSPECIFIED
+}
+
+func (x *DiscoverModelsRequest) GetApiKey() string {
+	if x != nil {
+		return x.ApiKey
+	}
+	return ""
+}
+
+func (x *DiscoverModelsRequest) GetBaseUrl() string {
+	if x != nil {
+		return x.BaseUrl
+	}
+	return ""
+}
+
+func (x *DiscoverModelsRequest) GetUseStoredApiKey() bool {
+	if x != nil {
+		return x.UseStoredApiKey
+	}
+	return false
+}
+
+type AvailableModel struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Id            string                 `protobuf:"bytes,1,opt,name=id,proto3" json:"id,omitempty"`
+	DisplayName   string                 `protobuf:"bytes,2,opt,name=display_name,json=displayName,proto3" json:"display_name,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *AvailableModel) Reset() {
+	*x = AvailableModel{}
+	mi := &file_chat_v1_chat_proto_msgTypes[4]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *AvailableModel) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*AvailableModel) ProtoMessage() {}
+
+func (x *AvailableModel) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[4]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use AvailableModel.ProtoReflect.Descriptor instead.
+func (*AvailableModel) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{4}
+}
+
+func (x *AvailableModel) GetId() string {
+	if x != nil {
+		return x.Id
+	}
+	return ""
+}
+
+func (x *AvailableModel) GetDisplayName() string {
+	if x != nil {
+		return x.DisplayName
+	}
+	return ""
+}
+
+type DiscoverModelsResponse struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Models        []*AvailableModel      `protobuf:"bytes,1,rep,name=models,proto3" json:"models,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *DiscoverModelsResponse) Reset() {
+	*x = DiscoverModelsResponse{}
+	mi := &file_chat_v1_chat_proto_msgTypes[5]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *DiscoverModelsResponse) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*DiscoverModelsResponse) ProtoMessage() {}
+
+func (x *DiscoverModelsResponse) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[5]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use DiscoverModelsResponse.ProtoReflect.Descriptor instead.
+func (*DiscoverModelsResponse) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{5}
+}
+
+func (x *DiscoverModelsResponse) GetModels() []*AvailableModel {
+	if x != nil {
+		return x.Models
+	}
+	return nil
+}
+
+type UpdateLLMConfigRequest struct {
+	state    protoimpl.MessageState `protogen:"open.v1"`
+	Harness  AgentHarness           `protobuf:"varint,1,opt,name=harness,proto3,enum=chat.v1.AgentHarness" json:"harness,omitempty"`
+	Provider LLMProvider            `protobuf:"varint,2,opt,name=provider,proto3,enum=chat.v1.LLMProvider" json:"provider,omitempty"`
+	// Write-only. An empty value preserves the existing key unless
+	// clear_api_key is true.
+	ApiKey       string `protobuf:"bytes,3,opt,name=api_key,json=apiKey,proto3" json:"api_key,omitempty"`
+	BaseUrl      string `protobuf:"bytes,4,opt,name=base_url,json=baseUrl,proto3" json:"base_url,omitempty"`
+	Model        string `protobuf:"bytes,5,opt,name=model,proto3" json:"model,omitempty"`
+	GooseBaseUrl string `protobuf:"bytes,7,opt,name=goose_base_url,json=gooseBaseUrl,proto3" json:"goose_base_url,omitempty"`
+	// Write-only shared secret for a remote Goose ACP endpoint.
+	GooseSecret      string `protobuf:"bytes,8,opt,name=goose_secret,json=gooseSecret,proto3" json:"goose_secret,omitempty"`
+	ClearApiKey      bool   `protobuf:"varint,9,opt,name=clear_api_key,json=clearApiKey,proto3" json:"clear_api_key,omitempty"`
+	ClearGooseSecret bool   `protobuf:"varint,10,opt,name=clear_goose_secret,json=clearGooseSecret,proto3" json:"clear_goose_secret,omitempty"`
+	unknownFields    protoimpl.UnknownFields
+	sizeCache        protoimpl.SizeCache
+}
+
+func (x *UpdateLLMConfigRequest) Reset() {
+	*x = UpdateLLMConfigRequest{}
+	mi := &file_chat_v1_chat_proto_msgTypes[6]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *UpdateLLMConfigRequest) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*UpdateLLMConfigRequest) ProtoMessage() {}
+
+func (x *UpdateLLMConfigRequest) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[6]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use UpdateLLMConfigRequest.ProtoReflect.Descriptor instead.
+func (*UpdateLLMConfigRequest) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{6}
+}
+
+func (x *UpdateLLMConfigRequest) GetHarness() AgentHarness {
+	if x != nil {
+		return x.Harness
+	}
+	return AgentHarness_AGENT_HARNESS_UNSPECIFIED
+}
+
+func (x *UpdateLLMConfigRequest) GetProvider() LLMProvider {
+	if x != nil {
+		return x.Provider
+	}
+	return LLMProvider_LLM_PROVIDER_UNSPECIFIED
+}
+
+func (x *UpdateLLMConfigRequest) GetApiKey() string {
+	if x != nil {
+		return x.ApiKey
+	}
+	return ""
+}
+
+func (x *UpdateLLMConfigRequest) GetBaseUrl() string {
+	if x != nil {
+		return x.BaseUrl
+	}
+	return ""
+}
+
+func (x *UpdateLLMConfigRequest) GetModel() string {
+	if x != nil {
+		return x.Model
+	}
+	return ""
+}
+
+func (x *UpdateLLMConfigRequest) GetGooseBaseUrl() string {
+	if x != nil {
+		return x.GooseBaseUrl
+	}
+	return ""
+}
+
+func (x *UpdateLLMConfigRequest) GetGooseSecret() string {
+	if x != nil {
+		return x.GooseSecret
+	}
+	return ""
+}
+
+func (x *UpdateLLMConfigRequest) GetClearApiKey() bool {
+	if x != nil {
+		return x.ClearApiKey
+	}
+	return false
+}
+
+func (x *UpdateLLMConfigRequest) GetClearGooseSecret() bool {
+	if x != nil {
+		return x.ClearGooseSecret
+	}
+	return false
+}
+
+type UpdateLLMConfigResponse struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Config        *LLMConfig             `protobuf:"bytes,1,opt,name=config,proto3" json:"config,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *UpdateLLMConfigResponse) Reset() {
+	*x = UpdateLLMConfigResponse{}
+	mi := &file_chat_v1_chat_proto_msgTypes[7]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *UpdateLLMConfigResponse) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*UpdateLLMConfigResponse) ProtoMessage() {}
+
+func (x *UpdateLLMConfigResponse) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[7]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use UpdateLLMConfigResponse.ProtoReflect.Descriptor instead.
+func (*UpdateLLMConfigResponse) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{7}
+}
+
+func (x *UpdateLLMConfigResponse) GetConfig() *LLMConfig {
+	if x != nil {
+		return x.Config
+	}
+	return nil
+}
+
+type ChatTurn struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Role          ChatRole               `protobuf:"varint,1,opt,name=role,proto3,enum=chat.v1.ChatRole" json:"role,omitempty"`
+	Content       string                 `protobuf:"bytes,2,opt,name=content,proto3" json:"content,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *ChatTurn) Reset() {
+	*x = ChatTurn{}
+	mi := &file_chat_v1_chat_proto_msgTypes[8]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ChatTurn) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ChatTurn) ProtoMessage() {}
+
+func (x *ChatTurn) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[8]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ChatTurn.ProtoReflect.Descriptor instead.
+func (*ChatTurn) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{8}
+}
+
+func (x *ChatTurn) GetRole() ChatRole {
+	if x != nil {
+		return x.Role
+	}
+	return ChatRole_CHAT_ROLE_UNSPECIFIED
+}
+
+func (x *ChatTurn) GetContent() string {
+	if x != nil {
+		return x.Content
+	}
+	return ""
+}
+
+type SendMessageRequest struct {
+	state          protoimpl.MessageState `protogen:"open.v1"`
+	ConversationId string                 `protobuf:"bytes,1,opt,name=conversation_id,json=conversationId,proto3" json:"conversation_id,omitempty"`
+	Content        string                 `protobuf:"bytes,2,opt,name=content,proto3" json:"content,omitempty"`
+	// Conversations remain client-side. Supplying prior turns keeps the server
+	// stateless while preserving multi-turn context.
+	History       []*ChatTurn `protobuf:"bytes,3,rep,name=history,proto3" json:"history,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *SendMessageRequest) Reset() {
+	*x = SendMessageRequest{}
+	mi := &file_chat_v1_chat_proto_msgTypes[9]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *SendMessageRequest) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*SendMessageRequest) ProtoMessage() {}
+
+func (x *SendMessageRequest) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[9]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use SendMessageRequest.ProtoReflect.Descriptor instead.
+func (*SendMessageRequest) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{9}
+}
+
+func (x *SendMessageRequest) GetConversationId() string {
+	if x != nil {
+		return x.ConversationId
+	}
+	return ""
+}
+
+func (x *SendMessageRequest) GetContent() string {
+	if x != nil {
+		return x.Content
+	}
+	return ""
+}
+
+func (x *SendMessageRequest) GetHistory() []*ChatTurn {
+	if x != nil {
+		return x.History
+	}
+	return nil
+}
+
+type TextDelta struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Content       string                 `protobuf:"bytes,1,opt,name=content,proto3" json:"content,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *TextDelta) Reset() {
+	*x = TextDelta{}
+	mi := &file_chat_v1_chat_proto_msgTypes[10]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *TextDelta) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*TextDelta) ProtoMessage() {}
+
+func (x *TextDelta) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[10]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use TextDelta.ProtoReflect.Descriptor instead.
+func (*TextDelta) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{10}
+}
+
+func (x *TextDelta) GetContent() string {
+	if x != nil {
+		return x.Content
+	}
+	return ""
+}
+
+type ToolCall struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Id            string                 `protobuf:"bytes,1,opt,name=id,proto3" json:"id,omitempty"`
+	Name          string                 `protobuf:"bytes,2,opt,name=name,proto3" json:"name,omitempty"`
+	Summary       string                 `protobuf:"bytes,3,opt,name=summary,proto3" json:"summary,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *ToolCall) Reset() {
+	*x = ToolCall{}
+	mi := &file_chat_v1_chat_proto_msgTypes[11]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ToolCall) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ToolCall) ProtoMessage() {}
+
+func (x *ToolCall) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[11]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ToolCall.ProtoReflect.Descriptor instead.
+func (*ToolCall) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{11}
+}
+
+func (x *ToolCall) GetId() string {
+	if x != nil {
+		return x.Id
+	}
+	return ""
+}
+
+func (x *ToolCall) GetName() string {
+	if x != nil {
+		return x.Name
+	}
+	return ""
+}
+
+func (x *ToolCall) GetSummary() string {
+	if x != nil {
+		return x.Summary
+	}
+	return ""
+}
+
+type ToolResult struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Id            string                 `protobuf:"bytes,1,opt,name=id,proto3" json:"id,omitempty"`
+	Name          string                 `protobuf:"bytes,2,opt,name=name,proto3" json:"name,omitempty"`
+	Success       bool                   `protobuf:"varint,3,opt,name=success,proto3" json:"success,omitempty"`
+	Summary       string                 `protobuf:"bytes,4,opt,name=summary,proto3" json:"summary,omitempty"`
+	Cancelled     bool                   `protobuf:"varint,5,opt,name=cancelled,proto3" json:"cancelled,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *ToolResult) Reset() {
+	*x = ToolResult{}
+	mi := &file_chat_v1_chat_proto_msgTypes[12]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ToolResult) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ToolResult) ProtoMessage() {}
+
+func (x *ToolResult) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[12]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ToolResult.ProtoReflect.Descriptor instead.
+func (*ToolResult) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{12}
+}
+
+func (x *ToolResult) GetId() string {
+	if x != nil {
+		return x.Id
+	}
+	return ""
+}
+
+func (x *ToolResult) GetName() string {
+	if x != nil {
+		return x.Name
+	}
+	return ""
+}
+
+func (x *ToolResult) GetSuccess() bool {
+	if x != nil {
+		return x.Success
+	}
+	return false
+}
+
+func (x *ToolResult) GetSummary() string {
+	if x != nil {
+		return x.Summary
+	}
+	return ""
+}
+
+func (x *ToolResult) GetCancelled() bool {
+	if x != nil {
+		return x.Cancelled
+	}
+	return false
+}
+
+type ToolConfirmationDetail struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	Label         string                 `protobuf:"bytes,1,opt,name=label,proto3" json:"label,omitempty"`
+	Value         string                 `protobuf:"bytes,2,opt,name=value,proto3" json:"value,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *ToolConfirmationDetail) Reset() {
+	*x = ToolConfirmationDetail{}
+	mi := &file_chat_v1_chat_proto_msgTypes[13]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ToolConfirmationDetail) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ToolConfirmationDetail) ProtoMessage() {}
+
+func (x *ToolConfirmationDetail) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[13]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ToolConfirmationDetail.ProtoReflect.Descriptor instead.
+func (*ToolConfirmationDetail) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{13}
+}
+
+func (x *ToolConfirmationDetail) GetLabel() string {
+	if x != nil {
+		return x.Label
+	}
+	return ""
+}
+
+func (x *ToolConfirmationDetail) GetValue() string {
+	if x != nil {
+		return x.Value
+	}
+	return ""
+}
+
+type ToolConfirmationRequired struct {
+	state protoimpl.MessageState `protogen:"open.v1"`
+	// Opaque, short-lived identifier bound to the current operator and org.
+	ConfirmationId string                    `protobuf:"bytes,1,opt,name=confirmation_id,json=confirmationId,proto3" json:"confirmation_id,omitempty"`
+	ToolCallId     string                    `protobuf:"bytes,2,opt,name=tool_call_id,json=toolCallId,proto3" json:"tool_call_id,omitempty"`
+	ToolName       string                    `protobuf:"bytes,3,opt,name=tool_name,json=toolName,proto3" json:"tool_name,omitempty"`
+	Title          string                    `protobuf:"bytes,4,opt,name=title,proto3" json:"title,omitempty"`
+	Description    string                    `protobuf:"bytes,5,opt,name=description,proto3" json:"description,omitempty"`
+	ConfirmLabel   string                    `protobuf:"bytes,6,opt,name=confirm_label,json=confirmLabel,proto3" json:"confirm_label,omitempty"`
+	Details        []*ToolConfirmationDetail `protobuf:"bytes,7,rep,name=details,proto3" json:"details,omitempty"`
+	unknownFields  protoimpl.UnknownFields
+	sizeCache      protoimpl.SizeCache
+}
+
+func (x *ToolConfirmationRequired) Reset() {
+	*x = ToolConfirmationRequired{}
+	mi := &file_chat_v1_chat_proto_msgTypes[14]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ToolConfirmationRequired) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ToolConfirmationRequired) ProtoMessage() {}
+
+func (x *ToolConfirmationRequired) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[14]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ToolConfirmationRequired.ProtoReflect.Descriptor instead.
+func (*ToolConfirmationRequired) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{14}
+}
+
+func (x *ToolConfirmationRequired) GetConfirmationId() string {
+	if x != nil {
+		return x.ConfirmationId
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetToolCallId() string {
+	if x != nil {
+		return x.ToolCallId
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetToolName() string {
+	if x != nil {
+		return x.ToolName
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetTitle() string {
+	if x != nil {
+		return x.Title
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetDescription() string {
+	if x != nil {
+		return x.Description
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetConfirmLabel() string {
+	if x != nil {
+		return x.ConfirmLabel
+	}
+	return ""
+}
+
+func (x *ToolConfirmationRequired) GetDetails() []*ToolConfirmationDetail {
+	if x != nil {
+		return x.Details
+	}
+	return nil
+}
+
+type ResolveToolConfirmationRequest struct {
+	state          protoimpl.MessageState   `protogen:"open.v1"`
+	ConfirmationId string                   `protobuf:"bytes,1,opt,name=confirmation_id,json=confirmationId,proto3" json:"confirmation_id,omitempty"`
+	Decision       ToolConfirmationDecision `protobuf:"varint,2,opt,name=decision,proto3,enum=chat.v1.ToolConfirmationDecision" json:"decision,omitempty"`
+	unknownFields  protoimpl.UnknownFields
+	sizeCache      protoimpl.SizeCache
+}
+
+func (x *ResolveToolConfirmationRequest) Reset() {
+	*x = ResolveToolConfirmationRequest{}
+	mi := &file_chat_v1_chat_proto_msgTypes[15]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ResolveToolConfirmationRequest) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ResolveToolConfirmationRequest) ProtoMessage() {}
+
+func (x *ResolveToolConfirmationRequest) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[15]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ResolveToolConfirmationRequest.ProtoReflect.Descriptor instead.
+func (*ResolveToolConfirmationRequest) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{15}
+}
+
+func (x *ResolveToolConfirmationRequest) GetConfirmationId() string {
+	if x != nil {
+		return x.ConfirmationId
+	}
+	return ""
+}
+
+func (x *ResolveToolConfirmationRequest) GetDecision() ToolConfirmationDecision {
+	if x != nil {
+		return x.Decision
+	}
+	return ToolConfirmationDecision_TOOL_CONFIRMATION_DECISION_UNSPECIFIED
+}
+
+type ResolveToolConfirmationResponse struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *ResolveToolConfirmationResponse) Reset() {
+	*x = ResolveToolConfirmationResponse{}
+	mi := &file_chat_v1_chat_proto_msgTypes[16]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *ResolveToolConfirmationResponse) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*ResolveToolConfirmationResponse) ProtoMessage() {}
+
+func (x *ResolveToolConfirmationResponse) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[16]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use ResolveToolConfirmationResponse.ProtoReflect.Descriptor instead.
+func (*ResolveToolConfirmationResponse) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{16}
+}
+
+type Done struct {
+	state         protoimpl.MessageState `protogen:"open.v1"`
+	FinishReason  string                 `protobuf:"bytes,1,opt,name=finish_reason,json=finishReason,proto3" json:"finish_reason,omitempty"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *Done) Reset() {
+	*x = Done{}
+	mi := &file_chat_v1_chat_proto_msgTypes[17]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *Done) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*Done) ProtoMessage() {}
+
+func (x *Done) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[17]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use Done.ProtoReflect.Descriptor instead.
+func (*Done) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{17}
+}
+
+func (x *Done) GetFinishReason() string {
+	if x != nil {
+		return x.FinishReason
+	}
+	return ""
+}
+
+type SendMessageResponse struct {
+	state     protoimpl.MessageState `protogen:"open.v1"`
+	MessageId string                 `protobuf:"bytes,1,opt,name=message_id,json=messageId,proto3" json:"message_id,omitempty"`
+	// Types that are valid to be assigned to Event:
+	//
+	//	*SendMessageResponse_TextDelta
+	//	*SendMessageResponse_ToolCall
+	//	*SendMessageResponse_ToolResult
+	//	*SendMessageResponse_Done
+	//	*SendMessageResponse_ConfirmationRequired
+	Event         isSendMessageResponse_Event `protobuf_oneof:"event"`
+	unknownFields protoimpl.UnknownFields
+	sizeCache     protoimpl.SizeCache
+}
+
+func (x *SendMessageResponse) Reset() {
+	*x = SendMessageResponse{}
+	mi := &file_chat_v1_chat_proto_msgTypes[18]
+	ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+	ms.StoreMessageInfo(mi)
+}
+
+func (x *SendMessageResponse) String() string {
+	return protoimpl.X.MessageStringOf(x)
+}
+
+func (*SendMessageResponse) ProtoMessage() {}
+
+func (x *SendMessageResponse) ProtoReflect() protoreflect.Message {
+	mi := &file_chat_v1_chat_proto_msgTypes[18]
+	if x != nil {
+		ms := protoimpl.X.MessageStateOf(protoimpl.Pointer(x))
+		if ms.LoadMessageInfo() == nil {
+			ms.StoreMessageInfo(mi)
+		}
+		return ms
+	}
+	return mi.MessageOf(x)
+}
+
+// Deprecated: Use SendMessageResponse.ProtoReflect.Descriptor instead.
+func (*SendMessageResponse) Descriptor() ([]byte, []int) {
+	return file_chat_v1_chat_proto_rawDescGZIP(), []int{18}
+}
+
+func (x *SendMessageResponse) GetMessageId() string {
+	if x != nil {
+		return x.MessageId
+	}
+	return ""
+}
+
+func (x *SendMessageResponse) GetEvent() isSendMessageResponse_Event {
+	if x != nil {
+		return x.Event
+	}
+	return nil
+}
+
+func (x *SendMessageResponse) GetTextDelta() *TextDelta {
+	if x != nil {
+		if x, ok := x.Event.(*SendMessageResponse_TextDelta); ok {
+			return x.TextDelta
+		}
+	}
+	return nil
+}
+
+func (x *SendMessageResponse) GetToolCall() *ToolCall {
+	if x != nil {
+		if x, ok := x.Event.(*SendMessageResponse_ToolCall); ok {
+			return x.ToolCall
+		}
+	}
+	return nil
+}
+
+func (x *SendMessageResponse) GetToolResult() *ToolResult {
+	if x != nil {
+		if x, ok := x.Event.(*SendMessageResponse_ToolResult); ok {
+			return x.ToolResult
+		}
+	}
+	return nil
+}
+
+func (x *SendMessageResponse) GetDone() *Done {
+	if x != nil {
+		if x, ok := x.Event.(*SendMessageResponse_Done); ok {
+			return x.Done
+		}
+	}
+	return nil
+}
+
+func (x *SendMessageResponse) GetConfirmationRequired() *ToolConfirmationRequired {
+	if x != nil {
+		if x, ok := x.Event.(*SendMessageResponse_ConfirmationRequired); ok {
+			return x.ConfirmationRequired
+		}
+	}
+	return nil
+}
+
+type isSendMessageResponse_Event interface {
+	isSendMessageResponse_Event()
+}
+
+type SendMessageResponse_TextDelta struct {
+	TextDelta *TextDelta `protobuf:"bytes,2,opt,name=text_delta,json=textDelta,proto3,oneof"`
+}
+
+type SendMessageResponse_ToolCall struct {
+	ToolCall *ToolCall `protobuf:"bytes,3,opt,name=tool_call,json=toolCall,proto3,oneof"`
+}
+
+type SendMessageResponse_ToolResult struct {
+	ToolResult *ToolResult `protobuf:"bytes,4,opt,name=tool_result,json=toolResult,proto3,oneof"`
+}
+
+type SendMessageResponse_Done struct {
+	Done *Done `protobuf:"bytes,5,opt,name=done,proto3,oneof"`
+}
+
+type SendMessageResponse_ConfirmationRequired struct {
+	ConfirmationRequired *ToolConfirmationRequired `protobuf:"bytes,6,opt,name=confirmation_required,json=confirmationRequired,proto3,oneof"`
+}
+
+func (*SendMessageResponse_TextDelta) isSendMessageResponse_Event() {}
+
+func (*SendMessageResponse_ToolCall) isSendMessageResponse_Event() {}
+
+func (*SendMessageResponse_ToolResult) isSendMessageResponse_Event() {}
+
+func (*SendMessageResponse_Done) isSendMessageResponse_Event() {}
+
+func (*SendMessageResponse_ConfirmationRequired) isSendMessageResponse_Event() {}
+
+var File_chat_v1_chat_proto protoreflect.FileDescriptor
+
+var file_chat_v1_chat_proto_rawDesc = string([]byte{
+	0x0a, 0x12, 0x63, 0x68, 0x61, 0x74, 0x2f, 0x76, 0x31, 0x2f, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x70,
+	0x72, 0x6f, 0x74, 0x6f, 0x12, 0x07, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x1a, 0x1b, 0x62,
+	0x75, 0x66, 0x2f, 0x76, 0x61, 0x6c, 0x69, 0x64, 0x61, 0x74, 0x65, 0x2f, 0x76, 0x61, 0x6c, 0x69,
+	0x64, 0x61, 0x74, 0x65, 0x2e, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x22, 0xb5, 0x02, 0x0a, 0x09, 0x4c,
+	0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x12, 0x2f, 0x0a, 0x07, 0x68, 0x61, 0x72, 0x6e,
+	0x65, 0x73, 0x73, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x15, 0x2e, 0x63, 0x68, 0x61, 0x74,
+	0x2e, 0x76, 0x31, 0x2e, 0x41, 0x67, 0x65, 0x6e, 0x74, 0x48, 0x61, 0x72, 0x6e, 0x65, 0x73, 0x73,
+	0x52, 0x07, 0x68, 0x61, 0x72, 0x6e, 0x65, 0x73, 0x73, 0x12, 0x30, 0x0a, 0x08, 0x70, 0x72, 0x6f,
+	0x76, 0x69, 0x64, 0x65, 0x72, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x14, 0x2e, 0x63, 0x68,
+	0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x4c, 0x4c, 0x4d, 0x50, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65,
+	0x72, 0x52, 0x08, 0x70, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72, 0x12, 0x1e, 0x0a, 0x0b, 0x68,
+	0x61, 0x73, 0x5f, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x18, 0x03, 0x20, 0x01, 0x28, 0x08,
+	0x52, 0x09, 0x68, 0x61, 0x73, 0x41, 0x70, 0x69, 0x4b, 0x65, 0x79, 0x12, 0x19, 0x0a, 0x08, 0x62,
+	0x61, 0x73, 0x65, 0x5f, 0x75, 0x72, 0x6c, 0x18, 0x04, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x62,
+	0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x12, 0x14, 0x0a, 0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x18,
+	0x05, 0x20, 0x01, 0x28, 0x09, 0x52, 0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x24, 0x0a, 0x0e,
+	0x67, 0x6f, 0x6f, 0x73, 0x65, 0x5f, 0x62, 0x61, 0x73, 0x65, 0x5f, 0x75, 0x72, 0x6c, 0x18, 0x07,
+	0x20, 0x01, 0x28, 0x09, 0x52, 0x0c, 0x67, 0x6f, 0x6f, 0x73, 0x65, 0x42, 0x61, 0x73, 0x65, 0x55,
+	0x72, 0x6c, 0x12, 0x28, 0x0a, 0x10, 0x68, 0x61, 0x73, 0x5f, 0x67, 0x6f, 0x6f, 0x73, 0x65, 0x5f,
+	0x73, 0x65, 0x63, 0x72, 0x65, 0x74, 0x18, 0x08, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0e, 0x68, 0x61,
+	0x73, 0x47, 0x6f, 0x6f, 0x73, 0x65, 0x53, 0x65, 0x63, 0x72, 0x65, 0x74, 0x12, 0x1e, 0x0a, 0x0a,
+	0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x75, 0x72, 0x65, 0x64, 0x18, 0x09, 0x20, 0x01, 0x28, 0x08,
+	0x52, 0x0a, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x75, 0x72, 0x65, 0x64, 0x4a, 0x04, 0x08, 0x06,
+	0x10, 0x07, 0x22, 0x15, 0x0a, 0x13, 0x47, 0x65, 0x74, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66,
+	0x69, 0x67, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x22, 0x42, 0x0a, 0x14, 0x47, 0x65, 0x74,
+	0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73,
+	0x65, 0x12, 0x2a, 0x0a, 0x06, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x18, 0x01, 0x20, 0x01, 0x28,
+	0x0b, 0x32, 0x12, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x4c, 0x4c, 0x4d, 0x43,
+	0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x06, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x22, 0xca, 0x01,
+	0x0a, 0x15, 0x44, 0x69, 0x73, 0x63, 0x6f, 0x76, 0x65, 0x72, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x73,
+	0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x3c, 0x0a, 0x08, 0x70, 0x72, 0x6f, 0x76, 0x69,
+	0x64, 0x65, 0x72, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x14, 0x2e, 0x63, 0x68, 0x61, 0x74,
+	0x2e, 0x76, 0x31, 0x2e, 0x4c, 0x4c, 0x4d, 0x50, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72, 0x42,
+	0x0a, 0xba, 0x48, 0x07, 0x82, 0x01, 0x04, 0x10, 0x01, 0x20, 0x00, 0x52, 0x08, 0x70, 0x72, 0x6f,
+	0x76, 0x69, 0x64, 0x65, 0x72, 0x12, 0x21, 0x0a, 0x07, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79,
+	0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x42, 0x08, 0xba, 0x48, 0x05, 0x72, 0x03, 0x18, 0x80, 0x40,
+	0x52, 0x06, 0x61, 0x70, 0x69, 0x4b, 0x65, 0x79, 0x12, 0x23, 0x0a, 0x08, 0x62, 0x61, 0x73, 0x65,
+	0x5f, 0x75, 0x72, 0x6c, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09, 0x42, 0x08, 0xba, 0x48, 0x05, 0x72,
+	0x03, 0x18, 0x80, 0x10, 0x52, 0x07, 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x12, 0x2b, 0x0a,
+	0x12, 0x75, 0x73, 0x65, 0x5f, 0x73, 0x74, 0x6f, 0x72, 0x65, 0x64, 0x5f, 0x61, 0x70, 0x69, 0x5f,
+	0x6b, 0x65, 0x79, 0x18, 0x04, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0f, 0x75, 0x73, 0x65, 0x53, 0x74,
+	0x6f, 0x72, 0x65, 0x64, 0x41, 0x70, 0x69, 0x4b, 0x65, 0x79, 0x22, 0x43, 0x0a, 0x0e, 0x41, 0x76,
+	0x61, 0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x0e, 0x0a, 0x02,
+	0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x02, 0x69, 0x64, 0x12, 0x21, 0x0a, 0x0c,
+	0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x5f, 0x6e, 0x61, 0x6d, 0x65, 0x18, 0x02, 0x20, 0x01,
+	0x28, 0x09, 0x52, 0x0b, 0x64, 0x69, 0x73, 0x70, 0x6c, 0x61, 0x79, 0x4e, 0x61, 0x6d, 0x65, 0x22,
+	0x49, 0x0a, 0x16, 0x44, 0x69, 0x73, 0x63, 0x6f, 0x76, 0x65, 0x72, 0x4d, 0x6f, 0x64, 0x65, 0x6c,
+	0x73, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x12, 0x2f, 0x0a, 0x06, 0x6d, 0x6f, 0x64,
+	0x65, 0x6c, 0x73, 0x18, 0x01, 0x20, 0x03, 0x28, 0x0b, 0x32, 0x17, 0x2e, 0x63, 0x68, 0x61, 0x74,
+	0x2e, 0x76, 0x31, 0x2e, 0x41, 0x76, 0x61, 0x69, 0x6c, 0x61, 0x62, 0x6c, 0x65, 0x4d, 0x6f, 0x64,
+	0x65, 0x6c, 0x52, 0x06, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x73, 0x22, 0xb2, 0x03, 0x0a, 0x16, 0x55,
+	0x70, 0x64, 0x61, 0x74, 0x65, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x65,
+	0x71, 0x75, 0x65, 0x73, 0x74, 0x12, 0x3b, 0x0a, 0x07, 0x68, 0x61, 0x72, 0x6e, 0x65, 0x73, 0x73,
+	0x18, 0x01, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x15, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31,
+	0x2e, 0x41, 0x67, 0x65, 0x6e, 0x74, 0x48, 0x61, 0x72, 0x6e, 0x65, 0x73, 0x73, 0x42, 0x0a, 0xba,
+	0x48, 0x07, 0x82, 0x01, 0x04, 0x10, 0x01, 0x20, 0x00, 0x52, 0x07, 0x68, 0x61, 0x72, 0x6e, 0x65,
+	0x73, 0x73, 0x12, 0x3c, 0x0a, 0x08, 0x70, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72, 0x18, 0x02,
+	0x20, 0x01, 0x28, 0x0e, 0x32, 0x14, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x4c,
+	0x4c, 0x4d, 0x50, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72, 0x42, 0x0a, 0xba, 0x48, 0x07, 0x82,
+	0x01, 0x04, 0x10, 0x01, 0x20, 0x00, 0x52, 0x08, 0x70, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72,
+	0x12, 0x21, 0x0a, 0x07, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x18, 0x03, 0x20, 0x01, 0x28,
+	0x09, 0x42, 0x08, 0xba, 0x48, 0x05, 0x72, 0x03, 0x18, 0x80, 0x40, 0x52, 0x06, 0x61, 0x70, 0x69,
+	0x4b, 0x65, 0x79, 0x12, 0x23, 0x0a, 0x08, 0x62, 0x61, 0x73, 0x65, 0x5f, 0x75, 0x72, 0x6c, 0x18,
+	0x04, 0x20, 0x01, 0x28, 0x09, 0x42, 0x08, 0xba, 0x48, 0x05, 0x72, 0x03, 0x18, 0x80, 0x10, 0x52,
+	0x07, 0x62, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x12, 0x20, 0x0a, 0x05, 0x6d, 0x6f, 0x64, 0x65,
+	0x6c, 0x18, 0x05, 0x20, 0x01, 0x28, 0x09, 0x42, 0x0a, 0xba, 0x48, 0x07, 0x72, 0x05, 0x10, 0x01,
+	0x18, 0xff, 0x01, 0x52, 0x05, 0x6d, 0x6f, 0x64, 0x65, 0x6c, 0x12, 0x2e, 0x0a, 0x0e, 0x67, 0x6f,
+	0x6f, 0x73, 0x65, 0x5f, 0x62, 0x61, 0x73, 0x65, 0x5f, 0x75, 0x72, 0x6c, 0x18, 0x07, 0x20, 0x01,
+	0x28, 0x09, 0x42, 0x08, 0xba, 0x48, 0x05, 0x72, 0x03, 0x18, 0x80, 0x10, 0x52, 0x0c, 0x67, 0x6f,
+	0x6f, 0x73, 0x65, 0x42, 0x61, 0x73, 0x65, 0x55, 0x72, 0x6c, 0x12, 0x2b, 0x0a, 0x0c, 0x67, 0x6f,
+	0x6f, 0x73, 0x65, 0x5f, 0x73, 0x65, 0x63, 0x72, 0x65, 0x74, 0x18, 0x08, 0x20, 0x01, 0x28, 0x09,
+	0x42, 0x08, 0xba, 0x48, 0x05, 0x72, 0x03, 0x18, 0x80, 0x40, 0x52, 0x0b, 0x67, 0x6f, 0x6f, 0x73,
+	0x65, 0x53, 0x65, 0x63, 0x72, 0x65, 0x74, 0x12, 0x22, 0x0a, 0x0d, 0x63, 0x6c, 0x65, 0x61, 0x72,
+	0x5f, 0x61, 0x70, 0x69, 0x5f, 0x6b, 0x65, 0x79, 0x18, 0x09, 0x20, 0x01, 0x28, 0x08, 0x52, 0x0b,
+	0x63, 0x6c, 0x65, 0x61, 0x72, 0x41, 0x70, 0x69, 0x4b, 0x65, 0x79, 0x12, 0x2c, 0x0a, 0x12, 0x63,
+	0x6c, 0x65, 0x61, 0x72, 0x5f, 0x67, 0x6f, 0x6f, 0x73, 0x65, 0x5f, 0x73, 0x65, 0x63, 0x72, 0x65,
+	0x74, 0x18, 0x0a, 0x20, 0x01, 0x28, 0x08, 0x52, 0x10, 0x63, 0x6c, 0x65, 0x61, 0x72, 0x47, 0x6f,
+	0x6f, 0x73, 0x65, 0x53, 0x65, 0x63, 0x72, 0x65, 0x74, 0x4a, 0x04, 0x08, 0x06, 0x10, 0x07, 0x22,
+	0x45, 0x0a, 0x17, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66,
+	0x69, 0x67, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x12, 0x2a, 0x0a, 0x06, 0x63, 0x6f,
+	0x6e, 0x66, 0x69, 0x67, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x12, 0x2e, 0x63, 0x68, 0x61,
+	0x74, 0x2e, 0x76, 0x31, 0x2e, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x06,
+	0x63, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x22, 0x64, 0x0a, 0x08, 0x43, 0x68, 0x61, 0x74, 0x54, 0x75,
+	0x72, 0x6e, 0x12, 0x31, 0x0a, 0x04, 0x72, 0x6f, 0x6c, 0x65, 0x18, 0x01, 0x20, 0x01, 0x28, 0x0e,
+	0x32, 0x11, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x43, 0x68, 0x61, 0x74, 0x52,
+	0x6f, 0x6c, 0x65, 0x42, 0x0a, 0xba, 0x48, 0x07, 0x82, 0x01, 0x04, 0x10, 0x01, 0x20, 0x00, 0x52,
+	0x04, 0x72, 0x6f, 0x6c, 0x65, 0x12, 0x25, 0x0a, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74,
+	0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x42, 0x0b, 0xba, 0x48, 0x08, 0x72, 0x06, 0x10, 0x01, 0x18,
+	0x80, 0x80, 0x02, 0x52, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0xa5, 0x01, 0x0a,
+	0x12, 0x53, 0x65, 0x6e, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x52, 0x65, 0x71, 0x75,
+	0x65, 0x73, 0x74, 0x12, 0x31, 0x0a, 0x0f, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61, 0x74,
+	0x69, 0x6f, 0x6e, 0x5f, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x42, 0x08, 0xba, 0x48,
+	0x05, 0x72, 0x03, 0x18, 0x80, 0x01, 0x52, 0x0e, 0x63, 0x6f, 0x6e, 0x76, 0x65, 0x72, 0x73, 0x61,
+	0x74, 0x69, 0x6f, 0x6e, 0x49, 0x64, 0x12, 0x25, 0x0a, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e,
+	0x74, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x42, 0x0b, 0xba, 0x48, 0x08, 0x72, 0x06, 0x10, 0x01,
+	0x18, 0x80, 0x80, 0x02, 0x52, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x12, 0x35, 0x0a,
+	0x07, 0x68, 0x69, 0x73, 0x74, 0x6f, 0x72, 0x79, 0x18, 0x03, 0x20, 0x03, 0x28, 0x0b, 0x32, 0x11,
+	0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x43, 0x68, 0x61, 0x74, 0x54, 0x75, 0x72,
+	0x6e, 0x42, 0x08, 0xba, 0x48, 0x05, 0x92, 0x01, 0x02, 0x10, 0x32, 0x52, 0x07, 0x68, 0x69, 0x73,
+	0x74, 0x6f, 0x72, 0x79, 0x22, 0x25, 0x0a, 0x09, 0x54, 0x65, 0x78, 0x74, 0x44, 0x65, 0x6c, 0x74,
+	0x61, 0x12, 0x18, 0x0a, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x18, 0x01, 0x20, 0x01,
+	0x28, 0x09, 0x52, 0x07, 0x63, 0x6f, 0x6e, 0x74, 0x65, 0x6e, 0x74, 0x22, 0x48, 0x0a, 0x08, 0x54,
+	0x6f, 0x6f, 0x6c, 0x43, 0x61, 0x6c, 0x6c, 0x12, 0x0e, 0x0a, 0x02, 0x69, 0x64, 0x18, 0x01, 0x20,
+	0x01, 0x28, 0x09, 0x52, 0x02, 0x69, 0x64, 0x12, 0x12, 0x0a, 0x04, 0x6e, 0x61, 0x6d, 0x65, 0x18,
+	0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x04, 0x6e, 0x61, 0x6d, 0x65, 0x12, 0x18, 0x0a, 0x07, 0x73,
+	0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09, 0x52, 0x07, 0x73, 0x75,
+	0x6d, 0x6d, 0x61, 0x72, 0x79, 0x22, 0x82, 0x01, 0x0a, 0x0a, 0x54, 0x6f, 0x6f, 0x6c, 0x52, 0x65,
+	0x73, 0x75, 0x6c, 0x74, 0x12, 0x0e, 0x0a, 0x02, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09,
+	0x52, 0x02, 0x69, 0x64, 0x12, 0x12, 0x0a, 0x04, 0x6e, 0x61, 0x6d, 0x65, 0x18, 0x02, 0x20, 0x01,
+	0x28, 0x09, 0x52, 0x04, 0x6e, 0x61, 0x6d, 0x65, 0x12, 0x18, 0x0a, 0x07, 0x73, 0x75, 0x63, 0x63,
+	0x65, 0x73, 0x73, 0x18, 0x03, 0x20, 0x01, 0x28, 0x08, 0x52, 0x07, 0x73, 0x75, 0x63, 0x63, 0x65,
+	0x73, 0x73, 0x12, 0x18, 0x0a, 0x07, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x18, 0x04, 0x20,
+	0x01, 0x28, 0x09, 0x52, 0x07, 0x73, 0x75, 0x6d, 0x6d, 0x61, 0x72, 0x79, 0x12, 0x1c, 0x0a, 0x09,
+	0x63, 0x61, 0x6e, 0x63, 0x65, 0x6c, 0x6c, 0x65, 0x64, 0x18, 0x05, 0x20, 0x01, 0x28, 0x08, 0x52,
+	0x09, 0x63, 0x61, 0x6e, 0x63, 0x65, 0x6c, 0x6c, 0x65, 0x64, 0x22, 0x44, 0x0a, 0x16, 0x54, 0x6f,
+	0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x44, 0x65,
+	0x74, 0x61, 0x69, 0x6c, 0x12, 0x14, 0x0a, 0x05, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x18, 0x01, 0x20,
+	0x01, 0x28, 0x09, 0x52, 0x05, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x12, 0x14, 0x0a, 0x05, 0x76, 0x61,
+	0x6c, 0x75, 0x65, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x05, 0x76, 0x61, 0x6c, 0x75, 0x65,
+	0x22, 0x9a, 0x02, 0x0a, 0x18, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d,
+	0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x12, 0x27, 0x0a,
+	0x0f, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x69, 0x64,
+	0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0e, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61,
+	0x74, 0x69, 0x6f, 0x6e, 0x49, 0x64, 0x12, 0x20, 0x0a, 0x0c, 0x74, 0x6f, 0x6f, 0x6c, 0x5f, 0x63,
+	0x61, 0x6c, 0x6c, 0x5f, 0x69, 0x64, 0x18, 0x02, 0x20, 0x01, 0x28, 0x09, 0x52, 0x0a, 0x74, 0x6f,
+	0x6f, 0x6c, 0x43, 0x61, 0x6c, 0x6c, 0x49, 0x64, 0x12, 0x1b, 0x0a, 0x09, 0x74, 0x6f, 0x6f, 0x6c,
+	0x5f, 0x6e, 0x61, 0x6d, 0x65, 0x18, 0x03, 0x20, 0x01, 0x28, 0x09, 0x52, 0x08, 0x74, 0x6f, 0x6f,
+	0x6c, 0x4e, 0x61, 0x6d, 0x65, 0x12, 0x14, 0x0a, 0x05, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x18, 0x04,
+	0x20, 0x01, 0x28, 0x09, 0x52, 0x05, 0x74, 0x69, 0x74, 0x6c, 0x65, 0x12, 0x20, 0x0a, 0x0b, 0x64,
+	0x65, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x18, 0x05, 0x20, 0x01, 0x28, 0x09,
+	0x52, 0x0b, 0x64, 0x65, 0x73, 0x63, 0x72, 0x69, 0x70, 0x74, 0x69, 0x6f, 0x6e, 0x12, 0x23, 0x0a,
+	0x0d, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x5f, 0x6c, 0x61, 0x62, 0x65, 0x6c, 0x18, 0x06,
+	0x20, 0x01, 0x28, 0x09, 0x52, 0x0c, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x4c, 0x61, 0x62,
+	0x65, 0x6c, 0x12, 0x39, 0x0a, 0x07, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x18, 0x07, 0x20,
+	0x03, 0x28, 0x0b, 0x32, 0x1f, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x54, 0x6f,
+	0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x44, 0x65,
+	0x74, 0x61, 0x69, 0x6c, 0x52, 0x07, 0x64, 0x65, 0x74, 0x61, 0x69, 0x6c, 0x73, 0x22, 0xa0, 0x01,
+	0x0a, 0x1e, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e,
+	0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74,
+	0x12, 0x33, 0x0a, 0x0f, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+	0x5f, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x42, 0x0a, 0xba, 0x48, 0x07, 0x72, 0x05,
+	0x10, 0x01, 0x18, 0x80, 0x01, 0x52, 0x0e, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74,
+	0x69, 0x6f, 0x6e, 0x49, 0x64, 0x12, 0x49, 0x0a, 0x08, 0x64, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f,
+	0x6e, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0e, 0x32, 0x21, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76,
+	0x31, 0x2e, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69,
+	0x6f, 0x6e, 0x44, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, 0x6e, 0x42, 0x0a, 0xba, 0x48, 0x07, 0x82,
+	0x01, 0x04, 0x10, 0x01, 0x20, 0x00, 0x52, 0x08, 0x64, 0x65, 0x63, 0x69, 0x73, 0x69, 0x6f, 0x6e,
+	0x22, 0x21, 0x0a, 0x1f, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x54, 0x6f, 0x6f, 0x6c, 0x43,
+	0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x73, 0x70, 0x6f,
+	0x6e, 0x73, 0x65, 0x22, 0x2b, 0x0a, 0x04, 0x44, 0x6f, 0x6e, 0x65, 0x12, 0x23, 0x0a, 0x0d, 0x66,
+	0x69, 0x6e, 0x69, 0x73, 0x68, 0x5f, 0x72, 0x65, 0x61, 0x73, 0x6f, 0x6e, 0x18, 0x01, 0x20, 0x01,
+	0x28, 0x09, 0x52, 0x0c, 0x66, 0x69, 0x6e, 0x69, 0x73, 0x68, 0x52, 0x65, 0x61, 0x73, 0x6f, 0x6e,
+	0x22, 0xdb, 0x02, 0x0a, 0x13, 0x53, 0x65, 0x6e, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65,
+	0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x12, 0x1d, 0x0a, 0x0a, 0x6d, 0x65, 0x73, 0x73,
+	0x61, 0x67, 0x65, 0x5f, 0x69, 0x64, 0x18, 0x01, 0x20, 0x01, 0x28, 0x09, 0x52, 0x09, 0x6d, 0x65,
+	0x73, 0x73, 0x61, 0x67, 0x65, 0x49, 0x64, 0x12, 0x33, 0x0a, 0x0a, 0x74, 0x65, 0x78, 0x74, 0x5f,
+	0x64, 0x65, 0x6c, 0x74, 0x61, 0x18, 0x02, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x12, 0x2e, 0x63, 0x68,
+	0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x54, 0x65, 0x78, 0x74, 0x44, 0x65, 0x6c, 0x74, 0x61, 0x48,
+	0x00, 0x52, 0x09, 0x74, 0x65, 0x78, 0x74, 0x44, 0x65, 0x6c, 0x74, 0x61, 0x12, 0x30, 0x0a, 0x09,
+	0x74, 0x6f, 0x6f, 0x6c, 0x5f, 0x63, 0x61, 0x6c, 0x6c, 0x18, 0x03, 0x20, 0x01, 0x28, 0x0b, 0x32,
+	0x11, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x61,
+	0x6c, 0x6c, 0x48, 0x00, 0x52, 0x08, 0x74, 0x6f, 0x6f, 0x6c, 0x43, 0x61, 0x6c, 0x6c, 0x12, 0x36,
+	0x0a, 0x0b, 0x74, 0x6f, 0x6f, 0x6c, 0x5f, 0x72, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x18, 0x04, 0x20,
+	0x01, 0x28, 0x0b, 0x32, 0x13, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x54, 0x6f,
+	0x6f, 0x6c, 0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x48, 0x00, 0x52, 0x0a, 0x74, 0x6f, 0x6f, 0x6c,
+	0x52, 0x65, 0x73, 0x75, 0x6c, 0x74, 0x12, 0x23, 0x0a, 0x04, 0x64, 0x6f, 0x6e, 0x65, 0x18, 0x05,
+	0x20, 0x01, 0x28, 0x0b, 0x32, 0x0d, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x44,
+	0x6f, 0x6e, 0x65, 0x48, 0x00, 0x52, 0x04, 0x64, 0x6f, 0x6e, 0x65, 0x12, 0x58, 0x0a, 0x15, 0x63,
+	0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x72, 0x65, 0x71, 0x75,
+	0x69, 0x72, 0x65, 0x64, 0x18, 0x06, 0x20, 0x01, 0x28, 0x0b, 0x32, 0x21, 0x2e, 0x63, 0x68, 0x61,
+	0x74, 0x2e, 0x76, 0x31, 0x2e, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d,
+	0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x71, 0x75, 0x69, 0x72, 0x65, 0x64, 0x48, 0x00, 0x52,
+	0x14, 0x63, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x71,
+	0x75, 0x69, 0x72, 0x65, 0x64, 0x42, 0x07, 0x0a, 0x05, 0x65, 0x76, 0x65, 0x6e, 0x74, 0x2a, 0x60,
+	0x0a, 0x0c, 0x41, 0x67, 0x65, 0x6e, 0x74, 0x48, 0x61, 0x72, 0x6e, 0x65, 0x73, 0x73, 0x12, 0x1d,
+	0x0a, 0x19, 0x41, 0x47, 0x45, 0x4e, 0x54, 0x5f, 0x48, 0x41, 0x52, 0x4e, 0x45, 0x53, 0x53, 0x5f,
+	0x55, 0x4e, 0x53, 0x50, 0x45, 0x43, 0x49, 0x46, 0x49, 0x45, 0x44, 0x10, 0x00, 0x12, 0x18, 0x0a,
+	0x14, 0x41, 0x47, 0x45, 0x4e, 0x54, 0x5f, 0x48, 0x41, 0x52, 0x4e, 0x45, 0x53, 0x53, 0x5f, 0x4e,
+	0x41, 0x54, 0x49, 0x56, 0x45, 0x10, 0x01, 0x12, 0x17, 0x0a, 0x13, 0x41, 0x47, 0x45, 0x4e, 0x54,
+	0x5f, 0x48, 0x41, 0x52, 0x4e, 0x45, 0x53, 0x53, 0x5f, 0x47, 0x4f, 0x4f, 0x53, 0x45, 0x10, 0x02,
+	0x2a, 0x92, 0x01, 0x0a, 0x0b, 0x4c, 0x4c, 0x4d, 0x50, 0x72, 0x6f, 0x76, 0x69, 0x64, 0x65, 0x72,
+	0x12, 0x1c, 0x0a, 0x18, 0x4c, 0x4c, 0x4d, 0x5f, 0x50, 0x52, 0x4f, 0x56, 0x49, 0x44, 0x45, 0x52,
+	0x5f, 0x55, 0x4e, 0x53, 0x50, 0x45, 0x43, 0x49, 0x46, 0x49, 0x45, 0x44, 0x10, 0x00, 0x12, 0x17,
+	0x0a, 0x13, 0x4c, 0x4c, 0x4d, 0x5f, 0x50, 0x52, 0x4f, 0x56, 0x49, 0x44, 0x45, 0x52, 0x5f, 0x4f,
+	0x50, 0x45, 0x4e, 0x41, 0x49, 0x10, 0x01, 0x12, 0x1a, 0x0a, 0x16, 0x4c, 0x4c, 0x4d, 0x5f, 0x50,
+	0x52, 0x4f, 0x56, 0x49, 0x44, 0x45, 0x52, 0x5f, 0x41, 0x4e, 0x54, 0x48, 0x52, 0x4f, 0x50, 0x49,
+	0x43, 0x10, 0x02, 0x12, 0x17, 0x0a, 0x13, 0x4c, 0x4c, 0x4d, 0x5f, 0x50, 0x52, 0x4f, 0x56, 0x49,
+	0x44, 0x45, 0x52, 0x5f, 0x4f, 0x4c, 0x4c, 0x41, 0x4d, 0x41, 0x10, 0x03, 0x12, 0x17, 0x0a, 0x13,
+	0x4c, 0x4c, 0x4d, 0x5f, 0x50, 0x52, 0x4f, 0x56, 0x49, 0x44, 0x45, 0x52, 0x5f, 0x43, 0x55, 0x53,
+	0x54, 0x4f, 0x4d, 0x10, 0x04, 0x2a, 0x52, 0x0a, 0x08, 0x43, 0x68, 0x61, 0x74, 0x52, 0x6f, 0x6c,
+	0x65, 0x12, 0x19, 0x0a, 0x15, 0x43, 0x48, 0x41, 0x54, 0x5f, 0x52, 0x4f, 0x4c, 0x45, 0x5f, 0x55,
+	0x4e, 0x53, 0x50, 0x45, 0x43, 0x49, 0x46, 0x49, 0x45, 0x44, 0x10, 0x00, 0x12, 0x12, 0x0a, 0x0e,
+	0x43, 0x48, 0x41, 0x54, 0x5f, 0x52, 0x4f, 0x4c, 0x45, 0x5f, 0x55, 0x53, 0x45, 0x52, 0x10, 0x01,
+	0x12, 0x17, 0x0a, 0x13, 0x43, 0x48, 0x41, 0x54, 0x5f, 0x52, 0x4f, 0x4c, 0x45, 0x5f, 0x41, 0x53,
+	0x53, 0x49, 0x53, 0x54, 0x41, 0x4e, 0x54, 0x10, 0x02, 0x2a, 0x95, 0x01, 0x0a, 0x18, 0x54, 0x6f,
+	0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x44, 0x65,
+	0x63, 0x69, 0x73, 0x69, 0x6f, 0x6e, 0x12, 0x2a, 0x0a, 0x26, 0x54, 0x4f, 0x4f, 0x4c, 0x5f, 0x43,
+	0x4f, 0x4e, 0x46, 0x49, 0x52, 0x4d, 0x41, 0x54, 0x49, 0x4f, 0x4e, 0x5f, 0x44, 0x45, 0x43, 0x49,
+	0x53, 0x49, 0x4f, 0x4e, 0x5f, 0x55, 0x4e, 0x53, 0x50, 0x45, 0x43, 0x49, 0x46, 0x49, 0x45, 0x44,
+	0x10, 0x00, 0x12, 0x26, 0x0a, 0x22, 0x54, 0x4f, 0x4f, 0x4c, 0x5f, 0x43, 0x4f, 0x4e, 0x46, 0x49,
+	0x52, 0x4d, 0x41, 0x54, 0x49, 0x4f, 0x4e, 0x5f, 0x44, 0x45, 0x43, 0x49, 0x53, 0x49, 0x4f, 0x4e,
+	0x5f, 0x41, 0x50, 0x50, 0x52, 0x4f, 0x56, 0x45, 0x10, 0x01, 0x12, 0x25, 0x0a, 0x21, 0x54, 0x4f,
+	0x4f, 0x4c, 0x5f, 0x43, 0x4f, 0x4e, 0x46, 0x49, 0x52, 0x4d, 0x41, 0x54, 0x49, 0x4f, 0x4e, 0x5f,
+	0x44, 0x45, 0x43, 0x49, 0x53, 0x49, 0x4f, 0x4e, 0x5f, 0x43, 0x41, 0x4e, 0x43, 0x45, 0x4c, 0x10,
+	0x02, 0x32, 0xbd, 0x03, 0x0a, 0x0b, 0x43, 0x68, 0x61, 0x74, 0x53, 0x65, 0x72, 0x76, 0x69, 0x63,
+	0x65, 0x12, 0x4b, 0x0a, 0x0c, 0x47, 0x65, 0x74, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69,
+	0x67, 0x12, 0x1c, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x47, 0x65, 0x74, 0x4c,
+	0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x1a,
+	0x1d, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x47, 0x65, 0x74, 0x4c, 0x4c, 0x4d,
+	0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x12, 0x51,
+	0x0a, 0x0e, 0x44, 0x69, 0x73, 0x63, 0x6f, 0x76, 0x65, 0x72, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x73,
+	0x12, 0x1e, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x44, 0x69, 0x73, 0x63, 0x6f,
+	0x76, 0x65, 0x72, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x73, 0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74,
+	0x1a, 0x1f, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x44, 0x69, 0x73, 0x63, 0x6f,
+	0x76, 0x65, 0x72, 0x4d, 0x6f, 0x64, 0x65, 0x6c, 0x73, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73,
+	0x65, 0x12, 0x54, 0x0a, 0x0f, 0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x4c, 0x4c, 0x4d, 0x43, 0x6f,
+	0x6e, 0x66, 0x69, 0x67, 0x12, 0x1f, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x55,
+	0x70, 0x64, 0x61, 0x74, 0x65, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52, 0x65,
+	0x71, 0x75, 0x65, 0x73, 0x74, 0x1a, 0x20, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e,
+	0x55, 0x70, 0x64, 0x61, 0x74, 0x65, 0x4c, 0x4c, 0x4d, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x67, 0x52,
+	0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73, 0x65, 0x12, 0x4a, 0x0a, 0x0b, 0x53, 0x65, 0x6e, 0x64, 0x4d,
+	0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x12, 0x1b, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31,
+	0x2e, 0x53, 0x65, 0x6e, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x52, 0x65, 0x71, 0x75,
+	0x65, 0x73, 0x74, 0x1a, 0x1c, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x53, 0x65,
+	0x6e, 0x64, 0x4d, 0x65, 0x73, 0x73, 0x61, 0x67, 0x65, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73,
+	0x65, 0x30, 0x01, 0x12, 0x6c, 0x0a, 0x17, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x54, 0x6f,
+	0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x12, 0x27,
+	0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76, 0x31, 0x2e, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65,
+	0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e, 0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e,
+	0x52, 0x65, 0x71, 0x75, 0x65, 0x73, 0x74, 0x1a, 0x28, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76,
+	0x31, 0x2e, 0x52, 0x65, 0x73, 0x6f, 0x6c, 0x76, 0x65, 0x54, 0x6f, 0x6f, 0x6c, 0x43, 0x6f, 0x6e,
+	0x66, 0x69, 0x72, 0x6d, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x52, 0x65, 0x73, 0x70, 0x6f, 0x6e, 0x73,
+	0x65, 0x42, 0x98, 0x01, 0x0a, 0x0b, 0x63, 0x6f, 0x6d, 0x2e, 0x63, 0x68, 0x61, 0x74, 0x2e, 0x76,
+	0x31, 0x42, 0x09, 0x43, 0x68, 0x61, 0x74, 0x50, 0x72, 0x6f, 0x74, 0x6f, 0x50, 0x01, 0x5a, 0x41,
+	0x67, 0x69, 0x74, 0x68, 0x75, 0x62, 0x2e, 0x63, 0x6f, 0x6d, 0x2f, 0x62, 0x6c, 0x6f, 0x63, 0x6b,
+	0x2f, 0x70, 0x72, 0x6f, 0x74, 0x6f, 0x2d, 0x66, 0x6c, 0x65, 0x65, 0x74, 0x2f, 0x73, 0x65, 0x72,
+	0x76, 0x65, 0x72, 0x2f, 0x67, 0x65, 0x6e, 0x65, 0x72, 0x61, 0x74, 0x65, 0x64, 0x2f, 0x67, 0x72,
+	0x70, 0x63, 0x2f, 0x63, 0x68, 0x61, 0x74, 0x2f, 0x76, 0x31, 0x3b, 0x63, 0x68, 0x61, 0x74, 0x76,
+	0x31, 0xa2, 0x02, 0x03, 0x43, 0x58, 0x58, 0xaa, 0x02, 0x07, 0x43, 0x68, 0x61, 0x74, 0x2e, 0x56,
+	0x31, 0xca, 0x02, 0x07, 0x43, 0x68, 0x61, 0x74, 0x5c, 0x56, 0x31, 0xe2, 0x02, 0x13, 0x43, 0x68,
+	0x61, 0x74, 0x5c, 0x56, 0x31, 0x5c, 0x47, 0x50, 0x42, 0x4d, 0x65, 0x74, 0x61, 0x64, 0x61, 0x74,
+	0x61, 0xea, 0x02, 0x08, 0x43, 0x68, 0x61, 0x74, 0x3a, 0x3a, 0x56, 0x31, 0x62, 0x06, 0x70, 0x72,
+	0x6f, 0x74, 0x6f, 0x33,
+})
+
+var (
+	file_chat_v1_chat_proto_rawDescOnce sync.Once
+	file_chat_v1_chat_proto_rawDescData []byte
+)
+
+func file_chat_v1_chat_proto_rawDescGZIP() []byte {
+	file_chat_v1_chat_proto_rawDescOnce.Do(func() {
+		file_chat_v1_chat_proto_rawDescData = protoimpl.X.CompressGZIP(unsafe.Slice(unsafe.StringData(file_chat_v1_chat_proto_rawDesc), len(file_chat_v1_chat_proto_rawDesc)))
+	})
+	return file_chat_v1_chat_proto_rawDescData
+}
+
+var file_chat_v1_chat_proto_enumTypes = make([]protoimpl.EnumInfo, 4)
+var file_chat_v1_chat_proto_msgTypes = make([]protoimpl.MessageInfo, 19)
+var file_chat_v1_chat_proto_goTypes = []any{
+	(AgentHarness)(0),                       // 0: chat.v1.AgentHarness
+	(LLMProvider)(0),                        // 1: chat.v1.LLMProvider
+	(ChatRole)(0),                           // 2: chat.v1.ChatRole
+	(ToolConfirmationDecision)(0),           // 3: chat.v1.ToolConfirmationDecision
+	(*LLMConfig)(nil),                       // 4: chat.v1.LLMConfig
+	(*GetLLMConfigRequest)(nil),             // 5: chat.v1.GetLLMConfigRequest
+	(*GetLLMConfigResponse)(nil),            // 6: chat.v1.GetLLMConfigResponse
+	(*DiscoverModelsRequest)(nil),           // 7: chat.v1.DiscoverModelsRequest
+	(*AvailableModel)(nil),                  // 8: chat.v1.AvailableModel
+	(*DiscoverModelsResponse)(nil),          // 9: chat.v1.DiscoverModelsResponse
+	(*UpdateLLMConfigRequest)(nil),          // 10: chat.v1.UpdateLLMConfigRequest
+	(*UpdateLLMConfigResponse)(nil),         // 11: chat.v1.UpdateLLMConfigResponse
+	(*ChatTurn)(nil),                        // 12: chat.v1.ChatTurn
+	(*SendMessageRequest)(nil),              // 13: chat.v1.SendMessageRequest
+	(*TextDelta)(nil),                       // 14: chat.v1.TextDelta
+	(*ToolCall)(nil),                        // 15: chat.v1.ToolCall
+	(*ToolResult)(nil),                      // 16: chat.v1.ToolResult
+	(*ToolConfirmationDetail)(nil),          // 17: chat.v1.ToolConfirmationDetail
+	(*ToolConfirmationRequired)(nil),        // 18: chat.v1.ToolConfirmationRequired
+	(*ResolveToolConfirmationRequest)(nil),  // 19: chat.v1.ResolveToolConfirmationRequest
+	(*ResolveToolConfirmationResponse)(nil), // 20: chat.v1.ResolveToolConfirmationResponse
+	(*Done)(nil),                            // 21: chat.v1.Done
+	(*SendMessageResponse)(nil),             // 22: chat.v1.SendMessageResponse
+}
+var file_chat_v1_chat_proto_depIdxs = []int32{
+	0,  // 0: chat.v1.LLMConfig.harness:type_name -> chat.v1.AgentHarness
+	1,  // 1: chat.v1.LLMConfig.provider:type_name -> chat.v1.LLMProvider
+	4,  // 2: chat.v1.GetLLMConfigResponse.config:type_name -> chat.v1.LLMConfig
+	1,  // 3: chat.v1.DiscoverModelsRequest.provider:type_name -> chat.v1.LLMProvider
+	8,  // 4: chat.v1.DiscoverModelsResponse.models:type_name -> chat.v1.AvailableModel
+	0,  // 5: chat.v1.UpdateLLMConfigRequest.harness:type_name -> chat.v1.AgentHarness
+	1,  // 6: chat.v1.UpdateLLMConfigRequest.provider:type_name -> chat.v1.LLMProvider
+	4,  // 7: chat.v1.UpdateLLMConfigResponse.config:type_name -> chat.v1.LLMConfig
+	2,  // 8: chat.v1.ChatTurn.role:type_name -> chat.v1.ChatRole
+	12, // 9: chat.v1.SendMessageRequest.history:type_name -> chat.v1.ChatTurn
+	17, // 10: chat.v1.ToolConfirmationRequired.details:type_name -> chat.v1.ToolConfirmationDetail
+	3,  // 11: chat.v1.ResolveToolConfirmationRequest.decision:type_name -> chat.v1.ToolConfirmationDecision
+	14, // 12: chat.v1.SendMessageResponse.text_delta:type_name -> chat.v1.TextDelta
+	15, // 13: chat.v1.SendMessageResponse.tool_call:type_name -> chat.v1.ToolCall
+	16, // 14: chat.v1.SendMessageResponse.tool_result:type_name -> chat.v1.ToolResult
+	21, // 15: chat.v1.SendMessageResponse.done:type_name -> chat.v1.Done
+	18, // 16: chat.v1.SendMessageResponse.confirmation_required:type_name -> chat.v1.ToolConfirmationRequired
+	5,  // 17: chat.v1.ChatService.GetLLMConfig:input_type -> chat.v1.GetLLMConfigRequest
+	7,  // 18: chat.v1.ChatService.DiscoverModels:input_type -> chat.v1.DiscoverModelsRequest
+	10, // 19: chat.v1.ChatService.UpdateLLMConfig:input_type -> chat.v1.UpdateLLMConfigRequest
+	13, // 20: chat.v1.ChatService.SendMessage:input_type -> chat.v1.SendMessageRequest
+	19, // 21: chat.v1.ChatService.ResolveToolConfirmation:input_type -> chat.v1.ResolveToolConfirmationRequest
+	6,  // 22: chat.v1.ChatService.GetLLMConfig:output_type -> chat.v1.GetLLMConfigResponse
+	9,  // 23: chat.v1.ChatService.DiscoverModels:output_type -> chat.v1.DiscoverModelsResponse
+	11, // 24: chat.v1.ChatService.UpdateLLMConfig:output_type -> chat.v1.UpdateLLMConfigResponse
+	22, // 25: chat.v1.ChatService.SendMessage:output_type -> chat.v1.SendMessageResponse
+	20, // 26: chat.v1.ChatService.ResolveToolConfirmation:output_type -> chat.v1.ResolveToolConfirmationResponse
+	22, // [22:27] is the sub-list for method output_type
+	17, // [17:22] is the sub-list for method input_type
+	17, // [17:17] is the sub-list for extension type_name
+	17, // [17:17] is the sub-list for extension extendee
+	0,  // [0:17] is the sub-list for field type_name
+}
+
+func init() { file_chat_v1_chat_proto_init() }
+func file_chat_v1_chat_proto_init() {
+	if File_chat_v1_chat_proto != nil {
+		return
+	}
+	file_chat_v1_chat_proto_msgTypes[18].OneofWrappers = []any{
+		(*SendMessageResponse_TextDelta)(nil),
+		(*SendMessageResponse_ToolCall)(nil),
+		(*SendMessageResponse_ToolResult)(nil),
+		(*SendMessageResponse_Done)(nil),
+		(*SendMessageResponse_ConfirmationRequired)(nil),
+	}
+	type x struct{}
+	out := protoimpl.TypeBuilder{
+		File: protoimpl.DescBuilder{
+			GoPackagePath: reflect.TypeOf(x{}).PkgPath(),
+			RawDescriptor: unsafe.Slice(unsafe.StringData(file_chat_v1_chat_proto_rawDesc), len(file_chat_v1_chat_proto_rawDesc)),
+			NumEnums:      4,
+			NumMessages:   19,
+			NumExtensions: 0,
+			NumServices:   1,
+		},
+		GoTypes:           file_chat_v1_chat_proto_goTypes,
+		DependencyIndexes: file_chat_v1_chat_proto_depIdxs,
+		EnumInfos:         file_chat_v1_chat_proto_enumTypes,
+		MessageInfos:      file_chat_v1_chat_proto_msgTypes,
+	}.Build()
+	File_chat_v1_chat_proto = out.File
+	file_chat_v1_chat_proto_goTypes = nil
+	file_chat_v1_chat_proto_depIdxs = nil
+}
diff --git a/server/generated/grpc/chat/v1/chatv1connect/chat.connect.go b/server/generated/grpc/chat/v1/chatv1connect/chat.connect.go
new file mode 100644
index 000000000..b4ea2db80
--- /dev/null
+++ b/server/generated/grpc/chat/v1/chatv1connect/chat.connect.go
@@ -0,0 +1,217 @@
+// Code generated by protoc-gen-connect-go. DO NOT EDIT.
+//
+// Source: chat/v1/chat.proto
+
+package chatv1connect
+
+import (
+	context "context"
+	errors "errors"
+	http "net/http"
+	strings "strings"
+
+	connect "connectrpc.com/connect"
+	v1 "github.com/block/proto-fleet/server/generated/grpc/chat/v1"
+)
+
+// This is a compile-time assertion to ensure that this generated file and the connect package are
+// compatible. If you get a compiler error that this constant is not defined, this code was
+// generated with a version of connect newer than the one compiled into your binary. You can fix the
+// problem by either regenerating this code with an older version of connect or updating the connect
+// version compiled into your binary.
+const _ = connect.IsAtLeastVersion0_1_0
+
+const (
+	// ChatServiceName is the fully-qualified name of the ChatService service.
+	ChatServiceName = "chat.v1.ChatService"
+)
+
+// These constants are the fully-qualified names of the RPCs defined in this package. They're
+// exposed at runtime as Spec.Procedure and as the final two segments of the HTTP route.
+//
+// Note that these are different from the fully-qualified method names used by
+// google.golang.org/protobuf/reflect/protoreflect. To convert from these constants to
+// reflection-formatted method names, remove the leading slash and convert the remaining slash to a
+// period.
+const (
+	// ChatServiceGetLLMConfigProcedure is the fully-qualified name of the ChatService's GetLLMConfig
+	// RPC.
+	ChatServiceGetLLMConfigProcedure = "/chat.v1.ChatService/GetLLMConfig"
+	// ChatServiceDiscoverModelsProcedure is the fully-qualified name of the ChatService's
+	// DiscoverModels RPC.
+	ChatServiceDiscoverModelsProcedure = "/chat.v1.ChatService/DiscoverModels"
+	// ChatServiceUpdateLLMConfigProcedure is the fully-qualified name of the ChatService's
+	// UpdateLLMConfig RPC.
+	ChatServiceUpdateLLMConfigProcedure = "/chat.v1.ChatService/UpdateLLMConfig"
+	// ChatServiceSendMessageProcedure is the fully-qualified name of the ChatService's SendMessage RPC.
+	ChatServiceSendMessageProcedure = "/chat.v1.ChatService/SendMessage"
+	// ChatServiceResolveToolConfirmationProcedure is the fully-qualified name of the ChatService's
+	// ResolveToolConfirmation RPC.
+	ChatServiceResolveToolConfirmationProcedure = "/chat.v1.ChatService/ResolveToolConfirmation"
+)
+
+// ChatServiceClient is a client for the chat.v1.ChatService service.
+type ChatServiceClient interface {
+	GetLLMConfig(context.Context, *connect.Request[v1.GetLLMConfigRequest]) (*connect.Response[v1.GetLLMConfigResponse], error)
+	DiscoverModels(context.Context, *connect.Request[v1.DiscoverModelsRequest]) (*connect.Response[v1.DiscoverModelsResponse], error)
+	UpdateLLMConfig(context.Context, *connect.Request[v1.UpdateLLMConfigRequest]) (*connect.Response[v1.UpdateLLMConfigResponse], error)
+	SendMessage(context.Context, *connect.Request[v1.SendMessageRequest]) (*connect.ServerStreamForClient[v1.SendMessageResponse], error)
+	// ResolveToolConfirmation approves or cancels a pending write requested by
+	// SendMessage. The original stream remains open and resumes after this RPC.
+	ResolveToolConfirmation(context.Context, *connect.Request[v1.ResolveToolConfirmationRequest]) (*connect.Response[v1.ResolveToolConfirmationResponse], error)
+}
+
+// NewChatServiceClient constructs a client for the chat.v1.ChatService service. By default, it uses
+// the Connect protocol with the binary Protobuf Codec, asks for gzipped responses, and sends
+// uncompressed requests. To use the gRPC or gRPC-Web protocols, supply the connect.WithGRPC() or
+// connect.WithGRPCWeb() options.
+//
+// The URL supplied here should be the base URL for the Connect or gRPC server (for example,
+// http://api.acme.com or https://acme.com/grpc).
+func NewChatServiceClient(httpClient connect.HTTPClient, baseURL string, opts ...connect.ClientOption) ChatServiceClient {
+	baseURL = strings.TrimRight(baseURL, "/")
+	return &chatServiceClient{
+		getLLMConfig: connect.NewClient[v1.GetLLMConfigRequest, v1.GetLLMConfigResponse](
+			httpClient,
+			baseURL+ChatServiceGetLLMConfigProcedure,
+			opts...,
+		),
+		discoverModels: connect.NewClient[v1.DiscoverModelsRequest, v1.DiscoverModelsResponse](
+			httpClient,
+			baseURL+ChatServiceDiscoverModelsProcedure,
+			opts...,
+		),
+		updateLLMConfig: connect.NewClient[v1.UpdateLLMConfigRequest, v1.UpdateLLMConfigResponse](
+			httpClient,
+			baseURL+ChatServiceUpdateLLMConfigProcedure,
+			opts...,
+		),
+		sendMessage: connect.NewClient[v1.SendMessageRequest, v1.SendMessageResponse](
+			httpClient,
+			baseURL+ChatServiceSendMessageProcedure,
+			opts...,
+		),
+		resolveToolConfirmation: connect.NewClient[v1.ResolveToolConfirmationRequest, v1.ResolveToolConfirmationResponse](
+			httpClient,
+			baseURL+ChatServiceResolveToolConfirmationProcedure,
+			opts...,
+		),
+	}
+}
+
+// chatServiceClient implements ChatServiceClient.
+type chatServiceClient struct {
+	getLLMConfig            *connect.Client[v1.GetLLMConfigRequest, v1.GetLLMConfigResponse]
+	discoverModels          *connect.Client[v1.DiscoverModelsRequest, v1.DiscoverModelsResponse]
+	updateLLMConfig         *connect.Client[v1.UpdateLLMConfigRequest, v1.UpdateLLMConfigResponse]
+	sendMessage             *connect.Client[v1.SendMessageRequest, v1.SendMessageResponse]
+	resolveToolConfirmation *connect.Client[v1.ResolveToolConfirmationRequest, v1.ResolveToolConfirmationResponse]
+}
+
+// GetLLMConfig calls chat.v1.ChatService.GetLLMConfig.
+func (c *chatServiceClient) GetLLMConfig(ctx context.Context, req *connect.Request[v1.GetLLMConfigRequest]) (*connect.Response[v1.GetLLMConfigResponse], error) {
+	return c.getLLMConfig.CallUnary(ctx, req)
+}
+
+// DiscoverModels calls chat.v1.ChatService.DiscoverModels.
+func (c *chatServiceClient) DiscoverModels(ctx context.Context, req *connect.Request[v1.DiscoverModelsRequest]) (*connect.Response[v1.DiscoverModelsResponse], error) {
+	return c.discoverModels.CallUnary(ctx, req)
+}
+
+// UpdateLLMConfig calls chat.v1.ChatService.UpdateLLMConfig.
+func (c *chatServiceClient) UpdateLLMConfig(ctx context.Context, req *connect.Request[v1.UpdateLLMConfigRequest]) (*connect.Response[v1.UpdateLLMConfigResponse], error) {
+	return c.updateLLMConfig.CallUnary(ctx, req)
+}
+
+// SendMessage calls chat.v1.ChatService.SendMessage.
+func (c *chatServiceClient) SendMessage(ctx context.Context, req *connect.Request[v1.SendMessageRequest]) (*connect.ServerStreamForClient[v1.SendMessageResponse], error) {
+	return c.sendMessage.CallServerStream(ctx, req)
+}
+
+// ResolveToolConfirmation calls chat.v1.ChatService.ResolveToolConfirmation.
+func (c *chatServiceClient) ResolveToolConfirmation(ctx context.Context, req *connect.Request[v1.ResolveToolConfirmationRequest]) (*connect.Response[v1.ResolveToolConfirmationResponse], error) {
+	return c.resolveToolConfirmation.CallUnary(ctx, req)
+}
+
+// ChatServiceHandler is an implementation of the chat.v1.ChatService service.
+type ChatServiceHandler interface {
+	GetLLMConfig(context.Context, *connect.Request[v1.GetLLMConfigRequest]) (*connect.Response[v1.GetLLMConfigResponse], error)
+	DiscoverModels(context.Context, *connect.Request[v1.DiscoverModelsRequest]) (*connect.Response[v1.DiscoverModelsResponse], error)
+	UpdateLLMConfig(context.Context, *connect.Request[v1.UpdateLLMConfigRequest]) (*connect.Response[v1.UpdateLLMConfigResponse], error)
+	SendMessage(context.Context, *connect.Request[v1.SendMessageRequest], *connect.ServerStream[v1.SendMessageResponse]) error
+	// ResolveToolConfirmation approves or cancels a pending write requested by
+	// SendMessage. The original stream remains open and resumes after this RPC.
+	ResolveToolConfirmation(context.Context, *connect.Request[v1.ResolveToolConfirmationRequest]) (*connect.Response[v1.ResolveToolConfirmationResponse], error)
+}
+
+// NewChatServiceHandler builds an HTTP handler from the service implementation. It returns the path
+// on which to mount the handler and the handler itself.
+//
+// By default, handlers support the Connect, gRPC, and gRPC-Web protocols with the binary Protobuf
+// and JSON codecs. They also support gzip compression.
+func NewChatServiceHandler(svc ChatServiceHandler, opts ...connect.HandlerOption) (string, http.Handler) {
+	chatServiceGetLLMConfigHandler := connect.NewUnaryHandler(
+		ChatServiceGetLLMConfigProcedure,
+		svc.GetLLMConfig,
+		opts...,
+	)
+	chatServiceDiscoverModelsHandler := connect.NewUnaryHandler(
+		ChatServiceDiscoverModelsProcedure,
+		svc.DiscoverModels,
+		opts...,
+	)
+	chatServiceUpdateLLMConfigHandler := connect.NewUnaryHandler(
+		ChatServiceUpdateLLMConfigProcedure,
+		svc.UpdateLLMConfig,
+		opts...,
+	)
+	chatServiceSendMessageHandler := connect.NewServerStreamHandler(
+		ChatServiceSendMessageProcedure,
+		svc.SendMessage,
+		opts...,
+	)
+	chatServiceResolveToolConfirmationHandler := connect.NewUnaryHandler(
+		ChatServiceResolveToolConfirmationProcedure,
+		svc.ResolveToolConfirmation,
+		opts...,
+	)
+	return "/chat.v1.ChatService/", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case ChatServiceGetLLMConfigProcedure:
+			chatServiceGetLLMConfigHandler.ServeHTTP(w, r)
+		case ChatServiceDiscoverModelsProcedure:
+			chatServiceDiscoverModelsHandler.ServeHTTP(w, r)
+		case ChatServiceUpdateLLMConfigProcedure:
+			chatServiceUpdateLLMConfigHandler.ServeHTTP(w, r)
+		case ChatServiceSendMessageProcedure:
+			chatServiceSendMessageHandler.ServeHTTP(w, r)
+		case ChatServiceResolveToolConfirmationProcedure:
+			chatServiceResolveToolConfirmationHandler.ServeHTTP(w, r)
+		default:
+			http.NotFound(w, r)
+		}
+	})
+}
+
+// UnimplementedChatServiceHandler returns CodeUnimplemented from all methods.
+type UnimplementedChatServiceHandler struct{}
+
+func (UnimplementedChatServiceHandler) GetLLMConfig(context.Context, *connect.Request[v1.GetLLMConfigRequest]) (*connect.Response[v1.GetLLMConfigResponse], error) {
+	return nil, connect.NewError(connect.CodeUnimplemented, errors.New("chat.v1.ChatService.GetLLMConfig is not implemented"))
+}
+
+func (UnimplementedChatServiceHandler) DiscoverModels(context.Context, *connect.Request[v1.DiscoverModelsRequest]) (*connect.Response[v1.DiscoverModelsResponse], error) {
+	return nil, connect.NewError(connect.CodeUnimplemented, errors.New("chat.v1.ChatService.DiscoverModels is not implemented"))
+}
+
+func (UnimplementedChatServiceHandler) UpdateLLMConfig(context.Context, *connect.Request[v1.UpdateLLMConfigRequest]) (*connect.Response[v1.UpdateLLMConfigResponse], error) {
+	return nil, connect.NewError(connect.CodeUnimplemented, errors.New("chat.v1.ChatService.UpdateLLMConfig is not implemented"))
+}
+
+func (UnimplementedChatServiceHandler) SendMessage(context.Context, *connect.Request[v1.SendMessageRequest], *connect.ServerStream[v1.SendMessageResponse]) error {
+	return connect.NewError(connect.CodeUnimplemented, errors.New("chat.v1.ChatService.SendMessage is not implemented"))
+}
+
+func (UnimplementedChatServiceHandler) ResolveToolConfirmation(context.Context, *connect.Request[v1.ResolveToolConfirmationRequest]) (*connect.Response[v1.ResolveToolConfirmationResponse], error) {
+	return nil, connect.NewError(connect.CodeUnimplemented, errors.New("chat.v1.ChatService.ResolveToolConfirmation is not implemented"))
+}
diff --git a/server/generated/sqlc/db.go b/server/generated/sqlc/db.go
index d84806ea8..66ca9b2d7 100644
--- a/server/generated/sqlc/db.go
+++ b/server/generated/sqlc/db.go
@@ -624,6 +624,9 @@ func Prepare(ctx context.Context, db DBTX) (*Queries, error) {
 	if q.getKnownSubnetsStmt, err = db.PrepareContext(ctx, getKnownSubnets); err != nil {
 		return nil, fmt.Errorf("error preparing query GetKnownSubnets: %w", err)
 	}
+	if q.getLLMConfigStmt, err = db.PrepareContext(ctx, getLLMConfig); err != nil {
+		return nil, fmt.Errorf("error preparing query GetLLMConfig: %w", err)
+	}
 	if q.getLatestAllDeviceMetricsStmt, err = db.PrepareContext(ctx, getLatestAllDeviceMetrics); err != nil {
 		return nil, fmt.Errorf("error preparing query GetLatestAllDeviceMetrics: %w", err)
 	}
@@ -1524,6 +1527,9 @@ func Prepare(ctx context.Context, db DBTX) (*Queries, error) {
 	if q.upsertFleetNodeSessionStmt, err = db.PrepareContext(ctx, upsertFleetNodeSession); err != nil {
 		return nil, fmt.Errorf("error preparing query UpsertFleetNodeSession: %w", err)
 	}
+	if q.upsertLLMConfigStmt, err = db.PrepareContext(ctx, upsertLLMConfig); err != nil {
+		return nil, fmt.Errorf("error preparing query UpsertLLMConfig: %w", err)
+	}
 	if q.upsertMQTTSourceStateStmt, err = db.PrepareContext(ctx, upsertMQTTSourceState); err != nil {
 		return nil, fmt.Errorf("error preparing query UpsertMQTTSourceState: %w", err)
 	}
@@ -2538,6 +2544,11 @@ func (q *Queries) Close() error {
 			err = fmt.Errorf("error closing getKnownSubnetsStmt: %w", cerr)
 		}
 	}
+	if q.getLLMConfigStmt != nil {
+		if cerr := q.getLLMConfigStmt.Close(); cerr != nil {
+			err = fmt.Errorf("error closing getLLMConfigStmt: %w", cerr)
+		}
+	}
 	if q.getLatestAllDeviceMetricsStmt != nil {
 		if cerr := q.getLatestAllDeviceMetricsStmt.Close(); cerr != nil {
 			err = fmt.Errorf("error closing getLatestAllDeviceMetricsStmt: %w", cerr)
@@ -4038,6 +4049,11 @@ func (q *Queries) Close() error {
 			err = fmt.Errorf("error closing upsertFleetNodeSessionStmt: %w", cerr)
 		}
 	}
+	if q.upsertLLMConfigStmt != nil {
+		if cerr := q.upsertLLMConfigStmt.Close(); cerr != nil {
+			err = fmt.Errorf("error closing upsertLLMConfigStmt: %w", cerr)
+		}
+	}
 	if q.upsertMQTTSourceStateStmt != nil {
 		if cerr := q.upsertMQTTSourceStateStmt.Close(); cerr != nil {
 			err = fmt.Errorf("error closing upsertMQTTSourceStateStmt: %w", cerr)
@@ -4292,6 +4308,7 @@ type Queries struct {
 	getInfrastructureControlSubnetsStmt                          *sql.Stmt
 	getInfrastructureDeviceStmt                                  *sql.Stmt
 	getKnownSubnetsStmt                                          *sql.Stmt
+	getLLMConfigStmt                                             *sql.Stmt
 	getLatestAllDeviceMetricsStmt                                *sql.Stmt
 	getLatestDeviceMetricsStmt                                   *sql.Stmt
 	getLatestFleetMetricRollupBucketStmt                         *sql.Stmt
@@ -4592,6 +4609,7 @@ type Queries struct {
 	upsertFleetMetricRollupsStmt                                 *sql.Stmt
 	upsertFleetNodeAuthChallengeStmt                             *sql.Stmt
 	upsertFleetNodeSessionStmt                                   *sql.Stmt
+	upsertLLMConfigStmt                                          *sql.Stmt
 	upsertMQTTSourceStateStmt                                    *sql.Stmt
 	upsertMinerCredentialsStmt                                   *sql.Stmt
 	upsertPermissionStmt                                         *sql.Stmt
@@ -4801,6 +4819,7 @@ func (q *Queries) WithTx(tx *sql.Tx) *Queries {
 		getInfrastructureControlSubnetsStmt:                          q.getInfrastructureControlSubnetsStmt,
 		getInfrastructureDeviceStmt:                                  q.getInfrastructureDeviceStmt,
 		getKnownSubnetsStmt:                                          q.getKnownSubnetsStmt,
+		getLLMConfigStmt:                                             q.getLLMConfigStmt,
 		getLatestAllDeviceMetricsStmt:                                q.getLatestAllDeviceMetricsStmt,
 		getLatestDeviceMetricsStmt:                                   q.getLatestDeviceMetricsStmt,
 		getLatestFleetMetricRollupBucketStmt:                         q.getLatestFleetMetricRollupBucketStmt,
@@ -5101,6 +5120,7 @@ func (q *Queries) WithTx(tx *sql.Tx) *Queries {
 		upsertFleetMetricRollupsStmt:                                 q.upsertFleetMetricRollupsStmt,
 		upsertFleetNodeAuthChallengeStmt:                             q.upsertFleetNodeAuthChallengeStmt,
 		upsertFleetNodeSessionStmt:                                   q.upsertFleetNodeSessionStmt,
+		upsertLLMConfigStmt:                                          q.upsertLLMConfigStmt,
 		upsertMQTTSourceStateStmt:                                    q.upsertMQTTSourceStateStmt,
 		upsertMinerCredentialsStmt:                                   q.upsertMinerCredentialsStmt,
 		upsertPermissionStmt:                                         q.upsertPermissionStmt,
diff --git a/server/generated/sqlc/llm_config.sql.go b/server/generated/sqlc/llm_config.sql.go
new file mode 100644
index 000000000..a7f34ff84
--- /dev/null
+++ b/server/generated/sqlc/llm_config.sql.go
@@ -0,0 +1,133 @@
+// Code generated by sqlc. DO NOT EDIT.
+// versions:
+//   sqlc v1.28.0
+// source: llm_config.sql
+
+package sqlc
+
+import (
+	"context"
+)
+
+const getLLMConfig = `-- name: GetLLMConfig :one
+SELECT
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted,
+    created_at,
+    updated_at
+FROM llm_config
+WHERE organization_id = $1
+`
+
+func (q *Queries) GetLLMConfig(ctx context.Context, organizationID int64) (LlmConfig, error) {
+	row := q.queryRow(ctx, q.getLLMConfigStmt, getLLMConfig, organizationID)
+	var i LlmConfig
+	err := row.Scan(
+		&i.OrganizationID,
+		&i.Harness,
+		&i.Provider,
+		&i.ApiKeyEncrypted,
+		&i.BaseUrl,
+		&i.Model,
+		&i.Temperature,
+		&i.GooseBaseUrl,
+		&i.GooseSecretEncrypted,
+		&i.CreatedAt,
+		&i.UpdatedAt,
+	)
+	return i, err
+}
+
+const upsertLLMConfig = `-- name: UpsertLLMConfig :one
+INSERT INTO llm_config (
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted
+) VALUES (
+    $1,
+    $2,
+    $3,
+    $4,
+    $5,
+    $6,
+    $7,
+    $8,
+    $9
+)
+ON CONFLICT (organization_id) DO UPDATE SET
+    harness = EXCLUDED.harness,
+    provider = EXCLUDED.provider,
+    api_key_encrypted = EXCLUDED.api_key_encrypted,
+    base_url = EXCLUDED.base_url,
+    model = EXCLUDED.model,
+    temperature = EXCLUDED.temperature,
+    goose_base_url = EXCLUDED.goose_base_url,
+    goose_secret_encrypted = EXCLUDED.goose_secret_encrypted,
+    updated_at = CURRENT_TIMESTAMP
+RETURNING
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted,
+    created_at,
+    updated_at
+`
+
+type UpsertLLMConfigParams struct {
+	OrganizationID       int64
+	Harness              string
+	Provider             string
+	ApiKeyEncrypted      string
+	BaseUrl              string
+	Model                string
+	Temperature          float64
+	GooseBaseUrl         string
+	GooseSecretEncrypted string
+}
+
+func (q *Queries) UpsertLLMConfig(ctx context.Context, arg UpsertLLMConfigParams) (LlmConfig, error) {
+	row := q.queryRow(ctx, q.upsertLLMConfigStmt, upsertLLMConfig,
+		arg.OrganizationID,
+		arg.Harness,
+		arg.Provider,
+		arg.ApiKeyEncrypted,
+		arg.BaseUrl,
+		arg.Model,
+		arg.Temperature,
+		arg.GooseBaseUrl,
+		arg.GooseSecretEncrypted,
+	)
+	var i LlmConfig
+	err := row.Scan(
+		&i.OrganizationID,
+		&i.Harness,
+		&i.Provider,
+		&i.ApiKeyEncrypted,
+		&i.BaseUrl,
+		&i.Model,
+		&i.Temperature,
+		&i.GooseBaseUrl,
+		&i.GooseSecretEncrypted,
+		&i.CreatedAt,
+		&i.UpdatedAt,
+	)
+	return i, err
+}
diff --git a/server/generated/sqlc/models.go b/server/generated/sqlc/models.go
index a8c2d5fea..1bb10f3d0 100644
--- a/server/generated/sqlc/models.go
+++ b/server/generated/sqlc/models.go
@@ -941,6 +941,23 @@ type InfrastructureDevice struct {
 	RackName     string
 }
 
+// Organization-scoped agent harness and encrypted BYOLLM provider configuration.
+type LlmConfig struct {
+	OrganizationID int64
+	Harness        string
+	Provider       string
+	// AES-GCM ciphertext produced by the fleet service master key; never returned over RPC.
+	ApiKeyEncrypted string
+	BaseUrl         string
+	Model           string
+	Temperature     float64
+	GooseBaseUrl    string
+	// Encrypted authentication secret for a future remote Goose ACP adapter.
+	GooseSecretEncrypted string
+	CreatedAt            time.Time
+	UpdatedAt            time.Time
+}
+
 type MinerCredential struct {
 	ID          int64
 	DeviceID    int64
diff --git a/server/generated/sqlc/querier.go b/server/generated/sqlc/querier.go
index 52ebe7c9a..dc93da136 100644
--- a/server/generated/sqlc/querier.go
+++ b/server/generated/sqlc/querier.go
@@ -599,6 +599,7 @@ type Querier interface {
 	GetInfrastructureControlSubnets(ctx context.Context, arg GetInfrastructureControlSubnetsParams) (string, error)
 	GetInfrastructureDevice(ctx context.Context, arg GetInfrastructureDeviceParams) (GetInfrastructureDeviceRow, error)
 	GetKnownSubnets(ctx context.Context, arg GetKnownSubnetsParams) ([]string, error)
+	GetLLMConfig(ctx context.Context, organizationID int64) (LlmConfig, error)
 	GetLatestAllDeviceMetrics(ctx context.Context, argTime time.Time) ([]DeviceMetric, error)
 	GetLatestDeviceMetrics(ctx context.Context, arg GetLatestDeviceMetricsParams) ([]DeviceMetric, error)
 	GetLatestFleetMetricRollupBucket(ctx context.Context) (time.Time, error)
@@ -1624,6 +1625,7 @@ type Querier interface {
 	UpsertFleetMetricRollups(ctx context.Context, arg UpsertFleetMetricRollupsParams) error
 	UpsertFleetNodeAuthChallenge(ctx context.Context, arg UpsertFleetNodeAuthChallengeParams) error
 	UpsertFleetNodeSession(ctx context.Context, arg UpsertFleetNodeSessionParams) error
+	UpsertLLMConfig(ctx context.Context, arg UpsertLLMConfigParams) (LlmConfig, error)
 	// Subscriber upserts source signal state on each successful message receive,
 	// after precedence/dedup processing. Singleton per source.
 	UpsertMQTTSourceState(ctx context.Context, arg UpsertMQTTSourceStateParams) error
diff --git a/server/generated/sqlc/retrying_querier.gen.go b/server/generated/sqlc/retrying_querier.gen.go
index 95fd11339..7000b38c7 100644
--- a/server/generated/sqlc/retrying_querier.gen.go
+++ b/server/generated/sqlc/retrying_querier.gen.go
@@ -2304,6 +2304,18 @@ func (q *retryingQuerier) GetKnownSubnets(ctx context.Context, arg GetKnownSubne
 	return result, err
 }
 
+func (q *retryingQuerier) GetLLMConfig(ctx context.Context, organizationID int64) (LlmConfig, error) {
+	var result LlmConfig
+	err := q.retrier.RetryQuery(ctx, "GetLLMConfig", func() error {
+		callResult, callErr := q.next.GetLLMConfig(ctx, organizationID)
+		if callErr == nil {
+			result = callResult
+		}
+		return callErr
+	})
+	return result, err
+}
+
 func (q *retryingQuerier) GetLatestAllDeviceMetrics(ctx context.Context, argTime time.Time) ([]DeviceMetric, error) {
 	var result []DeviceMetric
 	err := q.retrier.RetryQuery(ctx, "GetLatestAllDeviceMetrics", func() error {
@@ -5478,6 +5490,18 @@ func (q *retryingQuerier) UpsertFleetNodeSession(ctx context.Context, arg Upsert
 	})
 }
 
+func (q *retryingQuerier) UpsertLLMConfig(ctx context.Context, arg UpsertLLMConfigParams) (LlmConfig, error) {
+	var result LlmConfig
+	err := q.retrier.RetryQuery(ctx, "UpsertLLMConfig", func() error {
+		callResult, callErr := q.next.UpsertLLMConfig(ctx, arg)
+		if callErr == nil {
+			result = callResult
+		}
+		return callErr
+	})
+	return result, err
+}
+
 func (q *retryingQuerier) UpsertMQTTSourceState(ctx context.Context, arg UpsertMQTTSourceStateParams) error {
 	return q.retrier.RetryQuery(ctx, "UpsertMQTTSourceState", func() error {
 		return q.next.UpsertMQTTSourceState(ctx, arg)
diff --git a/server/internal/domain/chat/agent.go b/server/internal/domain/chat/agent.go
new file mode 100644
index 000000000..9d9598dac
--- /dev/null
+++ b/server/internal/domain/chat/agent.go
@@ -0,0 +1,358 @@
+package chat
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+
+	"github.com/block/proto-fleet/server/internal/domain/fleeterror"
+)
+
+const (
+	defaultMaxTurns     = 6
+	maxDeltaRunes       = 160
+	maxToolCallsPerTurn = 4
+	maxToolCallsPerRun  = 8
+)
+
+const systemPrompt = `You are Proto Fleet AI, an operations assistant for bitcoin-miner fleets.
+Use the available tools whenever a question depends on live fleet state. Never invent fleet data.
+You may request write actions with the available write tools. Every write is paused and shown to the operator for explicit confirmation before it executes. Never claim a change happened until its tool result reports success. If an action is cancelled, acknowledge the cancellation without immediately requesting it again.
+When a write tool requires device_identifiers, first resolve the target miners with resolve_miners unless the operator already supplied exact identifiers. Use limit 1000 when resolving all matching miners for a write. For miner command mutations, use preview_miner_action and execute_miner_action with selector.type "all_devices" for whole-fleet commands, selector.type "filter" for backend-resolvable subsets such as rack/site/building/status/model/IP filters, and selector.type "explicit" only when the operator supplied exact miner IDs or the target cannot be represented as a backend filter. If the destination rack is ambiguous, call list_racks before requesting the write. For rack slot placement, use list_racks to identify the rack layout and numbering origin, use get_rack_slots when existing occupancy matters, then call set_rack_slots with explicit 0-indexed row/column coordinates. For maintenance schedules, call preview_downtime_window before requesting create_downtime_window; explain that sleep schedules stop mining at the start time and do not automatically resume at end_time. Ask for clarification when miner, rack, slot, or schedule intent returns zero matches, ambiguous matches, truncated results, or cannot be converted safely.
+Answer directly and concisely. Mention what you checked only when it clarifies the scope or an access limitation.
+Format data for quick scanning:
+- Use short prose for a single fact, an explanation, or a recommendation.
+- Use a Markdown table for status breakdowns, comparisons, trends, or three or more related values or records.
+- Choose columns that match the question. Do not force unrelated facts into one table.
+- Put one entity or metric on each row, include units in headers when relevant, and right-align numeric columns with ---:.
+- Distinguish zero from None or Unavailable. Never infer missing values.
+- After a table, summarize the main conclusion in at most two sentences without repeating every value.`
+
+type Message struct {
+	Role       string
+	Content    string
+	ToolCalls  []ModelToolCall
+	ToolCallID string
+}
+
+type ModelToolCall struct {
+	ID        string
+	Name      string
+	Arguments json.RawMessage
+}
+
+type Completion struct {
+	Content   string
+	ToolCalls []ModelToolCall
+}
+
+type ToolDefinition struct {
+	Name                 string
+	Description          string
+	InputSchema          map[string]any
+	RequiresConfirmation bool
+}
+
+type ToolOutput struct {
+	Content string
+	Summary string
+}
+
+type ToolRegistry interface {
+	Definitions() []ToolDefinition
+	Execute(ctx context.Context, name string, arguments json.RawMessage) (ToolOutput, error)
+}
+
+type ToolConfirmationDetail struct {
+	Label string
+	Value string
+}
+
+type ToolConfirmation struct {
+	Title        string
+	Description  string
+	ConfirmLabel string
+	Details      []ToolConfirmationDetail
+}
+
+type ToolConfirmationProvider interface {
+	Confirmation(name string, arguments json.RawMessage) (*ToolConfirmation, error)
+}
+
+type ConfirmationDecision string
+
+const (
+	ConfirmationApproved  ConfirmationDecision = "approved"
+	ConfirmationCancelled ConfirmationDecision = "cancelled"
+)
+
+type ConfirmationRequest struct {
+	ToolCallID   string
+	ToolName     string
+	Confirmation ToolConfirmation
+}
+
+type ConfirmationGate interface {
+	Await(ctx context.Context, request ConfirmationRequest, notify func(confirmationID string) error) (ConfirmationDecision, error)
+}
+
+type ModelClient interface {
+	Complete(ctx context.Context, config RuntimeConfig, messages []Message, tools []ToolDefinition) (Completion, error)
+}
+
+type EventKind string
+
+const (
+	EventTextDelta            EventKind = "text_delta"
+	EventToolCall             EventKind = "tool_call"
+	EventToolResult           EventKind = "tool_result"
+	EventConfirmationRequired EventKind = "confirmation_required"
+	EventDone                 EventKind = "done"
+)
+
+type Event struct {
+	Kind           EventKind
+	Content        string
+	ToolCallID     string
+	ToolName       string
+	Summary        string
+	Success        bool
+	Cancelled      bool
+	ConfirmationID string
+	Confirmation   *ToolConfirmation
+}
+
+type Agent struct {
+	model         ModelClient
+	confirmations ConfirmationGate
+	maxTurns      int
+}
+
+type cachedToolResult struct {
+	output    ToolOutput
+	err       error
+	cancelled bool
+}
+
+func NewAgent(model ModelClient, confirmations ...ConfirmationGate) *Agent {
+	agent := &Agent{model: model, maxTurns: defaultMaxTurns}
+	if len(confirmations) > 0 {
+		agent.confirmations = confirmations[0]
+	}
+	return agent
+}
+
+func (a *Agent) Run(
+	ctx context.Context,
+	config RuntimeConfig,
+	history []Message,
+	content string,
+	tools ToolRegistry,
+	emit func(Event) error,
+) error {
+	if config.Harness == HarnessGoose {
+		return fleeterror.NewUnimplementedError("the Goose ACP harness is not available in this deployment; select the embedded harness")
+	}
+
+	messages := make([]Message, 0, len(history)+2)
+	messages = append(messages, Message{Role: "system", Content: systemPrompt})
+	messages = append(messages, history...)
+	messages = append(messages, Message{Role: "user", Content: content})
+	definitions := tools.Definitions()
+	requiresConfirmation := make(map[string]bool, len(definitions))
+	for _, definition := range definitions {
+		requiresConfirmation[definition.Name] = definition.RequiresConfirmation
+	}
+	toolCallsUsed := 0
+	toolResults := make(map[string]cachedToolResult)
+
+	for range a.maxTurns {
+		completion, err := a.model.Complete(ctx, config, messages, definitions)
+		if err != nil {
+			return err
+		}
+		if len(completion.ToolCalls) == 0 {
+			if strings.TrimSpace(completion.Content) == "" {
+				return fleeterror.NewUnavailableErrorf("the LLM provider returned an empty response")
+			}
+			for _, delta := range chunkText(completion.Content, maxDeltaRunes) {
+				if err := emit(Event{Kind: EventTextDelta, Content: delta}); err != nil {
+					return err
+				}
+			}
+			return emit(Event{Kind: EventDone, Summary: "stop", Success: true})
+		}
+		if len(completion.ToolCalls) > maxToolCallsPerTurn {
+			return fleeterror.NewFailedPreconditionErrorf(
+				"agent requested %d tool calls in one turn; limit is %d",
+				len(completion.ToolCalls),
+				maxToolCallsPerTurn,
+			)
+		}
+		if len(completion.ToolCalls) > maxToolCallsPerRun-toolCallsUsed {
+			return fleeterror.NewFailedPreconditionErrorf("agent exceeded the %d-call tool budget", maxToolCallsPerRun)
+		}
+		toolCallsUsed += len(completion.ToolCalls)
+
+		messages = append(messages, Message{
+			Role:      "assistant",
+			Content:   completion.Content,
+			ToolCalls: completion.ToolCalls,
+		})
+		for _, call := range completion.ToolCalls {
+			if err := emit(Event{
+				Kind:       EventToolCall,
+				ToolCallID: call.ID,
+				ToolName:   call.Name,
+				Summary:    toolCallSummary(call.Name),
+			}); err != nil {
+				return err
+			}
+
+			cacheKey := toolCallCacheKey(call)
+			cached, alreadyExecuted := toolResults[cacheKey]
+			if !alreadyExecuted {
+				if requiresConfirmation[call.Name] {
+					provider, ok := tools.(ToolConfirmationProvider)
+					if !ok || a.confirmations == nil {
+						return fleeterror.NewFailedPreconditionError("write tool confirmation is unavailable")
+					}
+					confirmation, confirmationErr := provider.Confirmation(call.Name, call.Arguments)
+					if confirmationErr != nil {
+						cached.err = confirmationErr
+					} else if confirmation == nil {
+						return fleeterror.NewFailedPreconditionErrorf("write tool %q did not provide confirmation details", call.Name)
+					} else {
+						decision, confirmationErr := a.confirmations.Await(ctx, ConfirmationRequest{
+							ToolCallID:   call.ID,
+							ToolName:     call.Name,
+							Confirmation: *confirmation,
+						}, func(confirmationID string) error {
+							return emit(Event{
+								Kind:           EventConfirmationRequired,
+								ToolCallID:     call.ID,
+								ToolName:       call.Name,
+								ConfirmationID: confirmationID,
+								Confirmation:   confirmation,
+							})
+						})
+						if confirmationErr != nil {
+							return confirmationErr
+						}
+						switch decision {
+						case ConfirmationCancelled:
+							cached.cancelled = true
+						case ConfirmationApproved:
+							cached.output, cached.err = tools.Execute(ctx, call.Name, call.Arguments)
+						default:
+							return fleeterror.NewFailedPreconditionError("invalid write tool confirmation decision")
+						}
+					}
+				} else {
+					cached.output, cached.err = tools.Execute(ctx, call.Name, call.Arguments)
+				}
+				toolResults[cacheKey] = cached
+			}
+			output, toolErr, cancelled := cached.output, cached.err, cached.cancelled
+			resultContent := output.Content
+			resultSummary := output.Summary
+			if cancelled {
+				resultContent = "Tool cancelled by operator"
+				resultSummary = "Cancelled by operator"
+			} else if toolErr != nil {
+				// Do not forward internal handler or authorization details to the
+				// external model provider. The operator receives the safe activity
+				// summary while the model only learns that the operation failed.
+				if requiresConfirmation[call.Name] {
+					resultContent = "Tool failed: the requested fleet change was not completed"
+					resultSummary = "Couldn't complete the requested change"
+				} else {
+					resultContent = "Tool failed: fleet data is unavailable or access was denied"
+					resultSummary = "Unable to read this fleet data"
+				}
+			}
+			if err := emit(Event{
+				Kind:       EventToolResult,
+				ToolCallID: call.ID,
+				ToolName:   call.Name,
+				Summary:    resultSummary,
+				Success:    toolErr == nil && !cancelled,
+				Cancelled:  cancelled,
+			}); err != nil {
+				return err
+			}
+			messages = append(messages, Message{
+				Role:       "tool",
+				Content:    resultContent,
+				ToolCallID: call.ID,
+			})
+		}
+	}
+
+	return fleeterror.NewFailedPreconditionErrorf("agent exceeded the %d-turn tool limit", a.maxTurns)
+}
+
+func toolCallCacheKey(call ModelToolCall) string {
+	arguments := call.Arguments
+	var normalized any
+	if json.Unmarshal(arguments, &normalized) == nil {
+		if encoded, err := json.Marshal(normalized); err == nil {
+			arguments = encoded
+		}
+	}
+	return call.Name + "\x00" + string(arguments)
+}
+
+func chunkText(content string, size int) []string {
+	runes := []rune(content)
+	chunks := make([]string, 0, (len(runes)+size-1)/size)
+	for len(runes) > 0 {
+		end := min(size, len(runes))
+		chunks = append(chunks, string(runes[:end]))
+		runes = runes[end:]
+	}
+	return chunks
+}
+
+func toolCallSummary(name string) string {
+	switch name {
+	case "get_miner_state_counts":
+		return "Checking fleet health"
+	case "get_site_health_summary":
+		return "Checking site health"
+	case "list_sites":
+		return "Reading site inventory"
+	case "list_pools":
+		return "Checking mining pools"
+	case "list_actionable_miner_issues":
+		return "Finding miner issues"
+	case "list_racks":
+		return "Reading rack inventory"
+	case "get_rack_slots":
+		return "Reading rack slots"
+	case "get_rack_health":
+		return "Checking rack health"
+	case "resolve_miners":
+		return "Resolving miners"
+	case "preview_miner_action":
+		return "Checking command support"
+	case "execute_miner_action":
+		return "Preparing miner command"
+	case "preview_downtime_window":
+		return "Previewing maintenance schedule"
+	case "create_downtime_window":
+		return "Preparing maintenance schedule"
+	case "create_site":
+		return "Preparing site creation"
+	case "create_rack":
+		return "Preparing rack creation"
+	case "move_miners_to_rack":
+		return "Preparing miner move"
+	case "set_rack_slots":
+		return "Preparing rack slot assignment"
+	case "clear_rack_slots":
+		return "Preparing rack slot clearing"
+	default:
+		return "Reading fleet data"
+	}
+}
diff --git a/server/internal/domain/chat/agent_test.go b/server/internal/domain/chat/agent_test.go
new file mode 100644
index 000000000..00bf6a7bc
--- /dev/null
+++ b/server/internal/domain/chat/agent_test.go
@@ -0,0 +1,235 @@
+package chat
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type sequenceModel struct {
+	completions []Completion
+	calls       int
+	messages    [][]Message
+}
+
+func (m *sequenceModel) Complete(_ context.Context, _ RuntimeConfig, messages []Message, _ []ToolDefinition) (Completion, error) {
+	m.messages = append(m.messages, append([]Message(nil), messages...))
+	completion := m.completions[m.calls]
+	m.calls++
+	return completion, nil
+}
+
+type recordingTools struct {
+	called bool
+	calls  int
+}
+
+type writeRecordingTools struct {
+	recordingTools
+}
+
+func (*writeRecordingTools) Definitions() []ToolDefinition {
+	return []ToolDefinition{{
+		Name:                 "create_site",
+		InputSchema:          map[string]any{"type": "object"},
+		RequiresConfirmation: true,
+	}}
+}
+
+func (*writeRecordingTools) Confirmation(string, json.RawMessage) (*ToolConfirmation, error) {
+	return &ToolConfirmation{
+		Title:        "Create this site?",
+		Description:  "Create a site.",
+		ConfirmLabel: "Create site",
+		Details:      []ToolConfirmationDetail{{Label: "Name", Value: "North"}},
+	}, nil
+}
+
+type staticConfirmationGate struct {
+	decision ConfirmationDecision
+	request  ConfirmationRequest
+}
+
+func (g *staticConfirmationGate) Await(_ context.Context, request ConfirmationRequest, notify func(string) error) (ConfirmationDecision, error) {
+	g.request = request
+	if err := notify("confirmation-1"); err != nil {
+		return "", err
+	}
+	return g.decision, nil
+}
+
+func (*recordingTools) Definitions() []ToolDefinition {
+	return []ToolDefinition{{Name: "list_sites", InputSchema: map[string]any{"type": "object"}}}
+}
+
+func (t *recordingTools) Execute(context.Context, string, json.RawMessage) (ToolOutput, error) {
+	t.called = true
+	t.calls++
+	return ToolOutput{Content: `{"sites":[]}`, Summary: "Read 0 sites"}, nil
+}
+
+func TestAgentRunsToolThenStreamsFinalAnswer(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{
+		{ToolCalls: []ModelToolCall{{ID: "call-1", Name: "list_sites", Arguments: json.RawMessage(`{}`)}}},
+		{Content: "There are no configured sites."},
+	}}
+	tools := &recordingTools{}
+	agent := NewAgent(model)
+	var events []Event
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "How many sites?", tools, func(event Event) error {
+		events = append(events, event)
+		return nil
+	})
+
+	require.NoError(t, err)
+	assert.True(t, tools.called)
+	require.NotEmpty(t, model.messages)
+	require.NotEmpty(t, model.messages[0])
+	assert.Contains(t, model.messages[0][0].Content, "Use a Markdown table for status breakdowns")
+	assert.Contains(t, model.messages[0][0].Content, "without repeating every value")
+	assert.Contains(t, model.messages[0][0].Content, "first resolve the target miners with resolve_miners")
+	assert.Equal(t, EventToolCall, events[0].Kind)
+	assert.Equal(t, EventToolResult, events[1].Kind)
+	assert.Equal(t, EventTextDelta, events[2].Kind)
+	assert.Equal(t, "There are no configured sites.", events[2].Content)
+	assert.Equal(t, EventDone, events[3].Kind)
+}
+
+func TestAgentRejectsUnavailableGooseHarness(t *testing.T) {
+	agent := NewAgent(&sequenceModel{})
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessGoose}, nil, "Hello", &recordingTools{}, func(Event) error { return nil })
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "Goose ACP harness")
+}
+
+func TestAgentRejectsTooManyToolCallsInOneTurnBeforeExecution(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{{ToolCalls: []ModelToolCall{
+		{ID: "call-1", Name: "list_sites", Arguments: json.RawMessage(`{}`)},
+		{ID: "call-2", Name: "list_sites", Arguments: json.RawMessage(`{}`)},
+		{ID: "call-3", Name: "list_sites", Arguments: json.RawMessage(`{}`)},
+		{ID: "call-4", Name: "list_sites", Arguments: json.RawMessage(`{}`)},
+		{ID: "call-5", Name: "list_sites", Arguments: json.RawMessage(`{}`)},
+	}}}}
+	tools := &recordingTools{}
+	agent := NewAgent(model)
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "List sites", tools, func(Event) error { return nil })
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "limit is 4")
+	assert.Zero(t, tools.calls)
+}
+
+func TestAgentEnforcesTotalToolCallBudgetBeforeExecutingOverflow(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{
+		{ToolCalls: []ModelToolCall{
+			{ID: "call-1", Name: "list_sites", Arguments: json.RawMessage(`{"request":1}`)},
+			{ID: "call-2", Name: "list_sites", Arguments: json.RawMessage(`{"request":2}`)},
+			{ID: "call-3", Name: "list_sites", Arguments: json.RawMessage(`{"request":3}`)},
+			{ID: "call-4", Name: "list_sites", Arguments: json.RawMessage(`{"request":4}`)},
+		}},
+		{ToolCalls: []ModelToolCall{
+			{ID: "call-5", Name: "list_sites", Arguments: json.RawMessage(`{"request":5}`)},
+			{ID: "call-6", Name: "list_sites", Arguments: json.RawMessage(`{"request":6}`)},
+			{ID: "call-7", Name: "list_sites", Arguments: json.RawMessage(`{"request":7}`)},
+			{ID: "call-8", Name: "list_sites", Arguments: json.RawMessage(`{"request":8}`)},
+		}},
+		{ToolCalls: []ModelToolCall{{ID: "call-9", Name: "list_sites", Arguments: json.RawMessage(`{"request":9}`)}}},
+	}}
+	tools := &recordingTools{}
+	agent := NewAgent(model)
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "Keep checking", tools, func(Event) error { return nil })
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "8-call tool budget")
+	assert.Equal(t, 8, tools.calls)
+}
+
+func TestAgentDeduplicatesEquivalentToolCalls(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{
+		{ToolCalls: []ModelToolCall{
+			{ID: "call-1", Name: "list_sites", Arguments: json.RawMessage(`{"b":2,"a":1}`)},
+			{ID: "call-2", Name: "list_sites", Arguments: json.RawMessage(`{"a":1,"b":2}`)},
+		}},
+		{Content: "Done."},
+	}}
+	tools := &recordingTools{}
+	agent := NewAgent(model)
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "List sites", tools, func(Event) error { return nil })
+
+	require.NoError(t, err)
+	assert.Equal(t, 1, tools.calls)
+	require.Len(t, model.messages, 2)
+	toolMessages := 0
+	for _, message := range model.messages[1] {
+		if message.Role == "tool" {
+			toolMessages++
+		}
+	}
+	assert.Equal(t, 2, toolMessages, "each provider tool-call id still receives a result")
+}
+
+func TestAgentWaitsForApprovalBeforeExecutingWriteTool(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{
+		{ToolCalls: []ModelToolCall{{ID: "call-1", Name: "create_site", Arguments: json.RawMessage(`{"name":"North"}`)}}},
+		{Content: "North was created."},
+	}}
+	tools := &writeRecordingTools{}
+	gate := &staticConfirmationGate{decision: ConfirmationApproved}
+	agent := NewAgent(model, gate)
+	var events []Event
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "Create North", tools, func(event Event) error {
+		if event.Kind == EventConfirmationRequired {
+			assert.False(t, tools.called, "write must remain paused while confirmation is emitted")
+		}
+		events = append(events, event)
+		return nil
+	})
+
+	require.NoError(t, err)
+	assert.True(t, tools.called)
+	assert.Equal(t, "call-1", gate.request.ToolCallID)
+	assert.Equal(t, "Create this site?", gate.request.Confirmation.Title)
+	require.Len(t, events, 5)
+	assert.Equal(t, EventToolCall, events[0].Kind)
+	assert.Equal(t, EventConfirmationRequired, events[1].Kind)
+	assert.Equal(t, "confirmation-1", events[1].ConfirmationID)
+	assert.Equal(t, EventToolResult, events[2].Kind)
+	assert.True(t, events[2].Success)
+	assert.Equal(t, EventTextDelta, events[3].Kind)
+	assert.Equal(t, EventDone, events[4].Kind)
+}
+
+func TestAgentCancellationDoesNotExecuteWriteTool(t *testing.T) {
+	model := &sequenceModel{completions: []Completion{
+		{ToolCalls: []ModelToolCall{{ID: "call-1", Name: "create_site", Arguments: json.RawMessage(`{"name":"North"}`)}}},
+		{Content: "I cancelled the site creation."},
+	}}
+	tools := &writeRecordingTools{}
+	gate := &staticConfirmationGate{decision: ConfirmationCancelled}
+	agent := NewAgent(model, gate)
+	var events []Event
+
+	err := agent.Run(t.Context(), RuntimeConfig{Harness: HarnessNative}, nil, "Create North", tools, func(event Event) error {
+		events = append(events, event)
+		return nil
+	})
+
+	require.NoError(t, err)
+	assert.False(t, tools.called)
+	require.Len(t, model.messages, 2)
+	assert.Equal(t, "Tool cancelled by operator", model.messages[1][3].Content)
+	require.Len(t, events, 5)
+	assert.Equal(t, EventToolResult, events[2].Kind)
+	assert.True(t, events[2].Cancelled)
+	assert.False(t, events[2].Success)
+}
diff --git a/server/internal/domain/chat/config.go b/server/internal/domain/chat/config.go
new file mode 100644
index 000000000..e892d0307
--- /dev/null
+++ b/server/internal/domain/chat/config.go
@@ -0,0 +1,400 @@
+package chat
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"net"
+	"net/url"
+	"strings"
+
+	"github.com/block/proto-fleet/server/internal/domain/fleeterror"
+)
+
+type Harness string
+
+const (
+	HarnessNative Harness = "native"
+	HarnessGoose  Harness = "goose"
+)
+
+type Provider string
+
+const (
+	ProviderUnspecified Provider = ""
+	ProviderOpenAI      Provider = "openai"
+	ProviderAnthropic   Provider = "anthropic"
+	ProviderOllama      Provider = "ollama"
+	ProviderCustom      Provider = "custom"
+)
+
+const DefaultTemperature = 0.2
+
+type ConfigRecord struct {
+	OrganizationID       int64
+	Harness              Harness
+	Provider             Provider
+	APIKeyEncrypted      string
+	BaseURL              string
+	Model                string
+	Temperature          float64
+	GooseBaseURL         string
+	GooseSecretEncrypted string
+}
+
+type ConfigView struct {
+	Harness        Harness
+	Provider       Provider
+	HasAPIKey      bool
+	BaseURL        string
+	Model          string
+	GooseBaseURL   string
+	HasGooseSecret bool
+	Configured     bool
+}
+
+type RuntimeConfig struct {
+	Harness      Harness
+	Provider     Provider
+	APIKey       string
+	BaseURL      string
+	Model        string
+	Temperature  float64
+	GooseBaseURL string
+	GooseSecret  string
+}
+
+type UpdateConfig struct {
+	Harness          Harness
+	Provider         Provider
+	APIKey           string
+	BaseURL          string
+	Model            string
+	GooseBaseURL     string
+	GooseSecret      string
+	ClearAPIKey      bool
+	ClearGooseSecret bool
+}
+
+type DiscoverConfig struct {
+	Provider        Provider
+	APIKey          string
+	BaseURL         string
+	UseStoredAPIKey bool
+}
+
+type ConfigStore interface {
+	Get(ctx context.Context, orgID int64) (ConfigRecord, error)
+	Upsert(ctx context.Context, record ConfigRecord) (ConfigRecord, error)
+}
+
+type Cipher interface {
+	Encrypt(plaintext []byte) (string, error)
+	Decrypt(ciphertext string) ([]byte, error)
+}
+
+type ConfigService struct {
+	store        ConfigStore
+	cipher       Cipher
+	egressConfig ProviderEgressConfig
+}
+
+func NewConfigService(store ConfigStore, cipher Cipher, egressConfig ProviderEgressConfig) *ConfigService {
+	return &ConfigService{store: store, cipher: cipher, egressConfig: egressConfig}
+}
+
+func defaultConfig() ConfigView {
+	return ConfigView{
+		Harness:  HarnessNative,
+		Provider: ProviderUnspecified,
+	}
+}
+
+func (s *ConfigService) Get(ctx context.Context, orgID int64) (ConfigView, error) {
+	if orgID == 0 {
+		return ConfigView{}, fleeterror.NewUnauthenticatedError("organization id missing on session")
+	}
+	record, err := s.store.Get(ctx, orgID)
+	if errors.Is(err, sql.ErrNoRows) {
+		return defaultConfig(), nil
+	}
+	if err != nil {
+		return ConfigView{}, err
+	}
+	return record.view(), nil
+}
+
+func (s *ConfigService) Update(ctx context.Context, orgID int64, input UpdateConfig) (ConfigView, error) {
+	if orgID == 0 {
+		return ConfigView{}, fleeterror.NewUnauthenticatedError("organization id missing on session")
+	}
+	if err := normalizeAndValidateConfig(&input, s.egressConfig); err != nil {
+		return ConfigView{}, err
+	}
+	if input.ClearAPIKey && input.APIKey != "" {
+		return ConfigView{}, fleeterror.NewInvalidArgumentError("api_key and clear_api_key cannot both be set")
+	}
+	if input.ClearGooseSecret && input.GooseSecret != "" {
+		return ConfigView{}, fleeterror.NewInvalidArgumentError("goose_secret and clear_goose_secret cannot both be set")
+	}
+
+	existing, err := s.store.Get(ctx, orgID)
+	if err != nil && !errors.Is(err, sql.ErrNoRows) {
+		return ConfigView{}, err
+	}
+	record := ConfigRecord{
+		OrganizationID:       orgID,
+		Harness:              input.Harness,
+		Provider:             input.Provider,
+		BaseURL:              input.BaseURL,
+		Model:                input.Model,
+		Temperature:          DefaultTemperature,
+		GooseBaseURL:         input.GooseBaseURL,
+		GooseSecretEncrypted: existing.GooseSecretEncrypted,
+	}
+	// Provider credentials are destination-bound. Carrying a redacted secret
+	// across a provider or endpoint change would let an update send the existing
+	// credential to a destination that never supplied it.
+	if existing.Provider == input.Provider && existing.BaseURL == input.BaseURL {
+		record.APIKeyEncrypted = existing.APIKeyEncrypted
+	}
+
+	if input.ClearAPIKey {
+		record.APIKeyEncrypted = ""
+	} else if input.APIKey != "" {
+		record.APIKeyEncrypted, err = s.cipher.Encrypt([]byte(input.APIKey))
+		if err != nil {
+			return ConfigView{}, fmt.Errorf("encrypt LLM API key: %w", err)
+		}
+	}
+	if input.ClearGooseSecret {
+		record.GooseSecretEncrypted = ""
+	} else if input.GooseSecret != "" {
+		record.GooseSecretEncrypted, err = s.cipher.Encrypt([]byte(input.GooseSecret))
+		if err != nil {
+			return ConfigView{}, fmt.Errorf("encrypt Goose secret: %w", err)
+		}
+	}
+
+	stored, err := s.store.Upsert(ctx, record)
+	if err != nil {
+		return ConfigView{}, err
+	}
+	return stored.view(), nil
+}
+
+func (s *ConfigService) Runtime(ctx context.Context, orgID int64) (RuntimeConfig, error) {
+	record, err := s.store.Get(ctx, orgID)
+	if errors.Is(err, sql.ErrNoRows) {
+		return RuntimeConfig{}, fleeterror.NewFailedPreconditionError("configure an AI provider in Settings > Agents before starting a chat")
+	}
+	if err != nil {
+		return RuntimeConfig{}, err
+	}
+	view := record.view()
+	if !view.Configured {
+		return RuntimeConfig{}, fleeterror.NewFailedPreconditionError("the selected AI provider configuration is incomplete")
+	}
+
+	runtime := RuntimeConfig{
+		Harness:      record.Harness,
+		Provider:     record.Provider,
+		BaseURL:      record.BaseURL,
+		Model:        record.Model,
+		Temperature:  record.Temperature,
+		GooseBaseURL: record.GooseBaseURL,
+	}
+	if record.APIKeyEncrypted != "" {
+		plaintext, err := s.cipher.Decrypt(record.APIKeyEncrypted)
+		if err != nil {
+			return RuntimeConfig{}, fmt.Errorf("decrypt LLM API key: %w", err)
+		}
+		runtime.APIKey = string(plaintext)
+	}
+	if record.GooseSecretEncrypted != "" {
+		plaintext, err := s.cipher.Decrypt(record.GooseSecretEncrypted)
+		if err != nil {
+			return RuntimeConfig{}, fmt.Errorf("decrypt Goose secret: %w", err)
+		}
+		runtime.GooseSecret = string(plaintext)
+	}
+	return runtime, nil
+}
+
+// DiscoveryRuntime resolves a one-time provider connection for model
+// discovery. A newly entered API key is used only for this request; an
+// operator can instead explicitly reuse the encrypted key already stored for
+// the same provider.
+func (s *ConfigService) DiscoveryRuntime(ctx context.Context, orgID int64, input DiscoverConfig) (RuntimeConfig, error) {
+	if orgID == 0 {
+		return RuntimeConfig{}, fleeterror.NewUnauthenticatedError("organization id missing on session")
+	}
+	if input.APIKey != "" && input.UseStoredAPIKey {
+		return RuntimeConfig{}, fleeterror.NewInvalidArgumentError("api_key and use_stored_api_key cannot both be set")
+	}
+	if err := normalizeAndValidateDiscoveryConfig(&input, s.egressConfig); err != nil {
+		return RuntimeConfig{}, err
+	}
+
+	runtime := RuntimeConfig{
+		Provider: input.Provider,
+		APIKey:   input.APIKey,
+		BaseURL:  input.BaseURL,
+	}
+	if input.UseStoredAPIKey {
+		record, err := s.store.Get(ctx, orgID)
+		if errors.Is(err, sql.ErrNoRows) {
+			return RuntimeConfig{}, fleeterror.NewFailedPreconditionError("no stored API key is available for the selected provider")
+		}
+		if err != nil {
+			return RuntimeConfig{}, err
+		}
+		if record.Provider != input.Provider || record.APIKeyEncrypted == "" {
+			return RuntimeConfig{}, fleeterror.NewFailedPreconditionError("no stored API key is available for the selected provider")
+		}
+		if record.BaseURL != input.BaseURL {
+			return RuntimeConfig{}, fleeterror.NewFailedPreconditionError("the stored API key can only be used with its saved provider base URL; enter the key again to use a different endpoint")
+		}
+		plaintext, err := s.cipher.Decrypt(record.APIKeyEncrypted)
+		if err != nil {
+			return RuntimeConfig{}, fmt.Errorf("decrypt LLM API key: %w", err)
+		}
+		runtime.APIKey = string(plaintext)
+	}
+	if input.Provider != ProviderOllama && runtime.APIKey == "" {
+		return RuntimeConfig{}, fleeterror.NewInvalidArgumentError("enter an API key or use the stored key before fetching models")
+	}
+	return runtime, nil
+}
+
+func (record ConfigRecord) view() ConfigView {
+	view := ConfigView{
+		Harness:        record.Harness,
+		Provider:       record.Provider,
+		HasAPIKey:      record.APIKeyEncrypted != "",
+		BaseURL:        record.BaseURL,
+		Model:          record.Model,
+		GooseBaseURL:   record.GooseBaseURL,
+		HasGooseSecret: record.GooseSecretEncrypted != "",
+	}
+	if record.Harness == HarnessNative {
+		view.Configured = record.Model != "" && (record.Provider == ProviderOllama || view.HasAPIKey)
+	}
+	// Goose is persisted now, but deliberately remains unavailable until the
+	// remote ACP authentication/session adapter lands.
+	return view
+}
+
+func normalizeAndValidateConfig(input *UpdateConfig, egressConfig ProviderEgressConfig) error {
+	input.BaseURL = strings.TrimRight(strings.TrimSpace(input.BaseURL), "/")
+	input.GooseBaseURL = strings.TrimRight(strings.TrimSpace(input.GooseBaseURL), "/")
+	input.Model = strings.TrimSpace(input.Model)
+
+	switch input.Harness {
+	case HarnessNative, HarnessGoose:
+	default:
+		return fleeterror.NewInvalidArgumentErrorf("unsupported agent harness %q", input.Harness)
+	}
+	switch input.Provider {
+	case ProviderUnspecified:
+		return fleeterror.NewInvalidArgumentError("select an LLM provider")
+	case ProviderOpenAI:
+		if input.BaseURL == "" {
+			input.BaseURL = "https://api.openai.com/v1"
+		}
+	case ProviderAnthropic:
+		if input.BaseURL == "" {
+			input.BaseURL = "https://api.anthropic.com"
+		}
+	case ProviderOllama:
+		if input.BaseURL == "" {
+			input.BaseURL = "http://127.0.0.1:11434"
+		}
+	case ProviderCustom:
+		if input.BaseURL == "" {
+			return fleeterror.NewInvalidArgumentError("a base URL is required for a custom provider")
+		}
+	default:
+		return fleeterror.NewInvalidArgumentErrorf("unsupported LLM provider %q", input.Provider)
+	}
+	if input.Model == "" {
+		return fleeterror.NewInvalidArgumentError("an LLM model is required")
+	}
+	if err := validateProviderEndpointURL(input.Provider, input.BaseURL, egressConfig); err != nil {
+		return err
+	}
+	if input.Harness == HarnessGoose {
+		if input.GooseBaseURL == "" {
+			return fleeterror.NewInvalidArgumentError("a Goose ACP base URL is required for the Goose harness")
+		}
+		if err := validateEndpointURL("Goose ACP base URL", input.GooseBaseURL); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func normalizeAndValidateDiscoveryConfig(input *DiscoverConfig, egressConfig ProviderEgressConfig) error {
+	input.BaseURL = strings.TrimRight(strings.TrimSpace(input.BaseURL), "/")
+	switch input.Provider {
+	case ProviderOpenAI:
+		if input.BaseURL == "" {
+			input.BaseURL = "https://api.openai.com/v1"
+		}
+	case ProviderAnthropic:
+		if input.BaseURL == "" {
+			input.BaseURL = "https://api.anthropic.com"
+		}
+	case ProviderOllama:
+		if input.BaseURL == "" {
+			input.BaseURL = "http://127.0.0.1:11434"
+		}
+	case ProviderCustom:
+		return fleeterror.NewInvalidArgumentError("model discovery is not available for custom providers")
+	case ProviderUnspecified:
+		return fleeterror.NewInvalidArgumentError("select an LLM provider")
+	default:
+		return fleeterror.NewInvalidArgumentErrorf("unsupported LLM provider %q", input.Provider)
+	}
+	return validateProviderEndpointURL(input.Provider, input.BaseURL, egressConfig)
+}
+
+func validateEndpointURL(label, raw string) error {
+	_, err := parseEndpointURL(label, raw)
+	return err
+}
+
+func parseEndpointURL(label, raw string) (*url.URL, error) {
+	parsed, err := url.Parse(raw)
+	if err != nil || parsed.Host == "" || (parsed.Scheme != "http" && parsed.Scheme != "https") {
+		return nil, fleeterror.NewInvalidArgumentErrorf("%s must be an absolute HTTP or HTTPS URL", label)
+	}
+	if parsed.User != nil || parsed.Fragment != "" || parsed.RawQuery != "" {
+		return nil, fleeterror.NewInvalidArgumentErrorf("%s cannot contain credentials, a query, or a fragment", label)
+	}
+	return parsed, nil
+}
+
+func validateProviderEndpointURL(provider Provider, raw string, egressConfig ProviderEgressConfig) error {
+	parsed, err := parseEndpointURL("provider base URL", raw)
+	if err != nil {
+		return err
+	}
+	if provider != ProviderOllama && parsed.Scheme != "https" {
+		return fleeterror.NewInvalidArgumentError("provider base URL must use HTTPS unless the provider is Ollama")
+	}
+
+	host := parsed.Hostname()
+	policy := providerEgressPolicyFor(provider, egressConfig)
+	if provider != ProviderOllama {
+		lowerHost := strings.ToLower(strings.TrimSuffix(host, "."))
+		if lowerHost == "localhost" || strings.HasSuffix(lowerHost, ".localhost") {
+			return fleeterror.NewInvalidArgumentError("provider base URL cannot target a private or internal address")
+		}
+	}
+	if ip := net.ParseIP(host); ip != nil && !providerIPAllowed(policy, ip) {
+		return fleeterror.NewInvalidArgumentError("provider base URL cannot target a private, internal, or reserved address")
+	}
+	return nil
+}
diff --git a/server/internal/domain/chat/config_test.go b/server/internal/domain/chat/config_test.go
new file mode 100644
index 000000000..70bac6033
--- /dev/null
+++ b/server/internal/domain/chat/config_test.go
@@ -0,0 +1,314 @@
+package chat
+
+import (
+	"context"
+	"database/sql"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+type memoryConfigStore struct {
+	record ConfigRecord
+	hasRow bool
+}
+
+func (s *memoryConfigStore) Get(context.Context, int64) (ConfigRecord, error) {
+	if !s.hasRow {
+		return ConfigRecord{}, sql.ErrNoRows
+	}
+	return s.record, nil
+}
+
+func (s *memoryConfigStore) Upsert(_ context.Context, record ConfigRecord) (ConfigRecord, error) {
+	s.record = record
+	s.hasRow = true
+	return record, nil
+}
+
+type prefixCipher struct{}
+
+func (prefixCipher) Encrypt(plaintext []byte) (string, error)  { return "enc:" + string(plaintext), nil }
+func (prefixCipher) Decrypt(ciphertext string) ([]byte, error) { return []byte(ciphertext[4:]), nil }
+
+func TestConfigServiceGetReturnsSafeDefaults(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	view, err := service.Get(t.Context(), 42)
+
+	require.NoError(t, err)
+	assert.Equal(t, HarnessNative, view.Harness)
+	assert.Equal(t, ProviderUnspecified, view.Provider)
+	assert.False(t, view.HasAPIKey)
+	assert.False(t, view.Configured)
+}
+
+func TestConfigServiceEncryptsAndNeverReturnsProviderSecret(t *testing.T) {
+	store := &memoryConfigStore{}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	view, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOpenAI,
+		APIKey:   "secret-key",
+		Model:    "model-name",
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, "enc:secret-key", store.record.APIKeyEncrypted)
+	assert.True(t, view.HasAPIKey)
+	assert.True(t, view.Configured)
+	assert.Equal(t, "https://api.openai.com/v1", view.BaseURL)
+	assert.Equal(t, DefaultTemperature, store.record.Temperature)
+
+	runtime, err := service.Runtime(t.Context(), 42)
+	require.NoError(t, err)
+	assert.Equal(t, "secret-key", runtime.APIKey)
+	assert.Equal(t, DefaultTemperature, runtime.Temperature)
+}
+
+func TestConfigServiceRequiresExplicitProviderSelection(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderUnspecified,
+		Model:    "model-name",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "select an LLM provider")
+}
+
+func TestConfigServiceBlankSecretPreservesExistingCiphertext(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Harness:         HarnessNative,
+		Provider:        ProviderOpenAI,
+		APIKeyEncrypted: "enc:old-key",
+		BaseURL:         "https://api.openai.com/v1",
+		Model:           "old-model",
+		Temperature:     0.3,
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOpenAI,
+		Model:    "new-model",
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, "enc:old-key", store.record.APIKeyEncrypted)
+	assert.Equal(t, "new-model", store.record.Model)
+}
+
+func TestConfigServiceDoesNotPreserveSecretForChangedProvider(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Harness:         HarnessNative,
+		Provider:        ProviderOpenAI,
+		APIKeyEncrypted: "enc:openai-key",
+		BaseURL:         "https://api.openai.com/v1",
+		Model:           "gpt-model",
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	view, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOllama,
+		Model:    "llama-model",
+	})
+
+	require.NoError(t, err)
+	assert.Empty(t, store.record.APIKeyEncrypted)
+	assert.False(t, view.HasAPIKey)
+}
+
+func TestConfigServiceDoesNotPreserveSecretForChangedBaseURL(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Harness:         HarnessNative,
+		Provider:        ProviderOpenAI,
+		APIKeyEncrypted: "enc:openai-key",
+		BaseURL:         "https://api.openai.com/v1",
+		Model:           "gpt-model",
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	view, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOpenAI,
+		BaseURL:  "https://llm-proxy.example.com/v1",
+		Model:    "gpt-model",
+	})
+
+	require.NoError(t, err)
+	assert.Empty(t, store.record.APIKeyEncrypted)
+	assert.False(t, view.HasAPIKey)
+	assert.False(t, view.Configured)
+}
+
+func TestConfigServiceRejectsCredentialBearingEndpoint(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderCustom,
+		BaseURL:  "https://user:password@example.com/v1",
+		Model:    "model-name",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "cannot contain credentials")
+}
+
+func TestConfigServiceRequiresHTTPSForCredentialBearingProviders(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOpenAI,
+		APIKey:   "secret-key",
+		BaseURL:  "http://api.example.com/v1",
+		Model:    "model-name",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "must use HTTPS")
+}
+
+func TestConfigServiceRejectsOllamaMetadataEndpoint(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOllama,
+		BaseURL:  "http://169.254.169.254",
+		Model:    "model-name",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "internal")
+}
+
+func TestConfigServiceRejectsPrivateOllamaEndpointByDefault(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOllama,
+		BaseURL:  "http://10.0.0.20:11434",
+		Model:    "model-name",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "private")
+}
+
+func TestConfigServiceAllowsPrivateOllamaEndpointWithDeploymentOptIn(t *testing.T) {
+	store := &memoryConfigStore{}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{AllowPrivateOllama: true})
+
+	view, err := service.Update(t.Context(), 42, UpdateConfig{
+		Harness:  HarnessNative,
+		Provider: ProviderOllama,
+		BaseURL:  "http://10.0.0.20:11434",
+		Model:    "model-name",
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, "http://10.0.0.20:11434", view.BaseURL)
+}
+
+func TestConfigServiceDiscoveryUsesOneTimeKeyAndProviderDefault(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	runtime, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{
+		Provider: ProviderOpenAI,
+		APIKey:   "one-time-key",
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, "one-time-key", runtime.APIKey)
+	assert.Equal(t, "https://api.openai.com/v1", runtime.BaseURL)
+}
+
+func TestConfigServiceDiscoveryCanUseStoredKeyForSameProvider(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Provider:        ProviderAnthropic,
+		APIKeyEncrypted: "enc:stored-key",
+		BaseURL:         "https://api.anthropic.com",
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	runtime, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{
+		Provider:        ProviderAnthropic,
+		UseStoredAPIKey: true,
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, "stored-key", runtime.APIKey)
+	assert.Equal(t, "https://api.anthropic.com", runtime.BaseURL)
+}
+
+func TestConfigServiceDiscoveryRejectsStoredKeyForChangedBaseURL(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Provider:        ProviderAnthropic,
+		APIKeyEncrypted: "enc:stored-key",
+		BaseURL:         "https://api.anthropic.com",
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{
+		Provider:        ProviderAnthropic,
+		BaseURL:         "https://llm-proxy.example.com",
+		UseStoredAPIKey: true,
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "saved provider base URL")
+}
+
+func TestConfigServiceDiscoveryRejectsStoredKeyFromAnotherProvider(t *testing.T) {
+	store := &memoryConfigStore{hasRow: true, record: ConfigRecord{
+		OrganizationID:  42,
+		Provider:        ProviderOpenAI,
+		APIKeyEncrypted: "enc:stored-key",
+	}}
+	service := NewConfigService(store, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{
+		Provider:        ProviderAnthropic,
+		UseStoredAPIKey: true,
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "no stored API key")
+}
+
+func TestConfigServiceDiscoveryAllowsOllamaWithoutKey(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	runtime, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{Provider: ProviderOllama})
+
+	require.NoError(t, err)
+	assert.Empty(t, runtime.APIKey)
+	assert.Equal(t, "http://127.0.0.1:11434", runtime.BaseURL)
+}
+
+func TestConfigServiceDiscoveryRejectsCustomProvider(t *testing.T) {
+	service := NewConfigService(&memoryConfigStore{}, prefixCipher{}, ProviderEgressConfig{})
+
+	_, err := service.DiscoveryRuntime(t.Context(), 42, DiscoverConfig{
+		Provider: ProviderCustom,
+		APIKey:   "one-time-key",
+		BaseURL:  "https://example.com/v1",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "not available for custom providers")
+}
diff --git a/server/internal/domain/chat/confirmation.go b/server/internal/domain/chat/confirmation.go
new file mode 100644
index 000000000..974c2c491
--- /dev/null
+++ b/server/internal/domain/chat/confirmation.go
@@ -0,0 +1,106 @@
+package chat
+
+import (
+	"context"
+	"fmt"
+	"sync"
+	"time"
+
+	"connectrpc.com/connect"
+	"github.com/google/uuid"
+
+	"github.com/block/proto-fleet/server/internal/domain/fleeterror"
+	"github.com/block/proto-fleet/server/internal/domain/session"
+)
+
+const defaultConfirmationTTL = 5 * time.Minute
+
+type pendingConfirmation struct {
+	organizationID int64
+	userID         int64
+	decision       chan ConfirmationDecision
+}
+
+// ConfirmationBroker coordinates a short-lived approval RPC with the
+// SendMessage stream that requested the write. Pending confirmations are
+// deliberately process-local and live only as long as the original stream.
+type ConfirmationBroker struct {
+	mu      sync.Mutex
+	pending map[string]*pendingConfirmation
+	ttl     time.Duration
+}
+
+func NewConfirmationBroker(ttl ...time.Duration) *ConfirmationBroker {
+	confirmationTTL := defaultConfirmationTTL
+	if len(ttl) > 0 {
+		confirmationTTL = ttl[0]
+	}
+	return &ConfirmationBroker{
+		pending: make(map[string]*pendingConfirmation),
+		ttl:     confirmationTTL,
+	}
+}
+
+func (b *ConfirmationBroker) Await(
+	ctx context.Context,
+	_ ConfirmationRequest,
+	notify func(confirmationID string) error,
+) (ConfirmationDecision, error) {
+	info, err := session.GetInfo(ctx)
+	if err != nil {
+		return "", fleeterror.NewUnauthenticatedError("authentication required")
+	}
+	confirmationID := uuid.NewString()
+	pending := &pendingConfirmation{
+		organizationID: info.OrganizationID,
+		userID:         info.UserID,
+		decision:       make(chan ConfirmationDecision, 1),
+	}
+
+	b.mu.Lock()
+	b.pending[confirmationID] = pending
+	b.mu.Unlock()
+	defer func() {
+		b.mu.Lock()
+		delete(b.pending, confirmationID)
+		b.mu.Unlock()
+	}()
+
+	if err := notify(confirmationID); err != nil {
+		return "", err
+	}
+
+	timer := time.NewTimer(b.ttl)
+	defer timer.Stop()
+	select {
+	case decision := <-pending.decision:
+		return decision, nil
+	case <-timer.C:
+		return "", fleeterror.NewPlainError("tool confirmation expired", connect.CodeDeadlineExceeded)
+	case <-ctx.Done():
+		return "", fmt.Errorf("wait for tool confirmation: %w", ctx.Err())
+	}
+}
+
+func (b *ConfirmationBroker) Resolve(ctx context.Context, confirmationID string, decision ConfirmationDecision) error {
+	info, err := session.GetInfo(ctx)
+	if err != nil {
+		return fleeterror.NewUnauthenticatedError("authentication required")
+	}
+	if decision != ConfirmationApproved && decision != ConfirmationCancelled {
+		return fleeterror.NewInvalidArgumentError("invalid tool confirmation decision")
+	}
+
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	pending, ok := b.pending[confirmationID]
+	if !ok || pending.organizationID != info.OrganizationID || pending.userID != info.UserID {
+		return fleeterror.NewNotFoundError("tool confirmation not found or expired")
+	}
+	select {
+	case pending.decision <- decision:
+		return nil
+	default:
+		return fleeterror.NewFailedPreconditionError("tool confirmation was already resolved")
+	}
+}
diff --git a/server/internal/domain/chat/confirmation_test.go b/server/internal/domain/chat/confirmation_test.go
new file mode 100644
index 000000000..4d564e425
--- /dev/null
+++ b/server/internal/domain/chat/confirmation_test.go
@@ -0,0 +1,85 @@
+package chat
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"connectrpc.com/authn"
+	"connectrpc.com/connect"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/block/proto-fleet/server/internal/domain/fleeterror"
+	"github.com/block/proto-fleet/server/internal/domain/session"
+)
+
+type confirmationResult struct {
+	decision ConfirmationDecision
+	err      error
+}
+
+func confirmationTestContext(ctx context.Context, userID, orgID int64) context.Context {
+	return authn.SetInfo(ctx, &session.Info{UserID: userID, OrganizationID: orgID})
+}
+
+func TestConfirmationBrokerResumesAwaitingRequestForSameOperator(t *testing.T) {
+	broker := NewConfirmationBroker(time.Second)
+	ctx := confirmationTestContext(t.Context(), 7, 42)
+	confirmationID := make(chan string, 1)
+	result := make(chan confirmationResult, 1)
+
+	go func() {
+		decision, err := broker.Await(ctx, ConfirmationRequest{}, func(id string) error {
+			confirmationID <- id
+			return nil
+		})
+		result <- confirmationResult{decision: decision, err: err}
+	}()
+
+	id := <-confirmationID
+	require.NoError(t, broker.Resolve(ctx, id, ConfirmationApproved))
+	resolved := <-result
+	require.NoError(t, resolved.err)
+	assert.Equal(t, ConfirmationApproved, resolved.decision)
+}
+
+func TestConfirmationBrokerDoesNotExposePendingRequestAcrossOperators(t *testing.T) {
+	broker := NewConfirmationBroker(time.Second)
+	ownerCtx := confirmationTestContext(t.Context(), 7, 42)
+	otherCtx := confirmationTestContext(t.Context(), 8, 42)
+	confirmationID := make(chan string, 1)
+	result := make(chan confirmationResult, 1)
+
+	go func() {
+		decision, err := broker.Await(ownerCtx, ConfirmationRequest{}, func(id string) error {
+			confirmationID <- id
+			return nil
+		})
+		result <- confirmationResult{decision: decision, err: err}
+	}()
+
+	id := <-confirmationID
+	err := broker.Resolve(otherCtx, id, ConfirmationApproved)
+	require.Error(t, err)
+	var fleetErr fleeterror.FleetError
+	require.ErrorAs(t, err, &fleetErr)
+	assert.Equal(t, connect.CodeNotFound, fleetErr.GRPCCode)
+	require.NoError(t, broker.Resolve(ownerCtx, id, ConfirmationCancelled))
+	resolved := <-result
+	require.NoError(t, resolved.err)
+	assert.Equal(t, ConfirmationCancelled, resolved.decision)
+}
+
+func TestConfirmationBrokerExpiresUnresolvedRequest(t *testing.T) {
+	broker := NewConfirmationBroker(time.Millisecond)
+	ctx := confirmationTestContext(context.Background(), 7, 42)
+
+	decision, err := broker.Await(ctx, ConfirmationRequest{}, func(string) error { return nil })
+
+	require.Error(t, err)
+	assert.Empty(t, decision)
+	var fleetErr fleeterror.FleetError
+	require.ErrorAs(t, err, &fleetErr)
+	assert.Equal(t, connect.CodeDeadlineExceeded, fleetErr.GRPCCode)
+}
diff --git a/server/internal/domain/chat/egress.go b/server/internal/domain/chat/egress.go
new file mode 100644
index 000000000..2ec5aad88
--- /dev/null
+++ b/server/internal/domain/chat/egress.go
@@ -0,0 +1,119 @@
+package chat
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+	"net/http"
+	"time"
+)
+
+const providerRequestTimeout = 90 * time.Second
+
+var errProviderDestinationDisallowed = errors.New("provider destination resolves to a disallowed internal address")
+
+type providerEgressPolicy struct {
+	allowLoopback bool
+	allowPrivate  bool
+}
+
+// ProviderEgressConfig contains deployment-owner-controlled exceptions to the
+// default provider destination policy. Keeping this separate from the
+// organization-scoped LLM settings prevents an API user from granting their
+// own access to services on the Fleet server's private network.
+type ProviderEgressConfig struct {
+	AllowPrivateOllama bool `help:"Allow Ollama to connect to RFC1918/private addresses in addition to loopback. Enable only when the deployment owner trusts that network." default:"false" env:"ALLOW_PRIVATE_OLLAMA"`
+}
+
+func providerEgressPolicyFor(provider Provider, config ProviderEgressConfig) providerEgressPolicy {
+	if provider != ProviderOllama {
+		return providerEgressPolicy{}
+	}
+	return providerEgressPolicy{
+		allowLoopback: true,
+		allowPrivate:  config.AllowPrivateOllama,
+	}
+}
+
+// newProviderHTTPClient validates the address that is actually dialed so a
+// hostname cannot pass configuration validation and later rebind to an
+// internal service. Redirects and environment proxies are disabled because
+// either would move resolution outside this transport's destination policy.
+func newProviderHTTPClient(policy providerEgressPolicy) *http.Client {
+	dialer := &net.Dialer{Timeout: providerRequestTimeout}
+	transport, _ := http.DefaultTransport.(*http.Transport)
+	transport = transport.Clone()
+	transport.Proxy = nil
+	transport.DialContext = func(ctx context.Context, network, addr string) (net.Conn, error) {
+		host, port, err := net.SplitHostPort(addr)
+		if err != nil {
+			return nil, fmt.Errorf("split provider destination: %w", err)
+		}
+		ips, err := net.DefaultResolver.LookupIP(ctx, "ip", host)
+		if err != nil {
+			return nil, fmt.Errorf("resolve provider destination: %w", err)
+		}
+		lastErr := errors.New("provider destination has no dialable address")
+		for _, ip := range ips {
+			if !providerIPAllowed(policy, ip) {
+				lastErr = errProviderDestinationDisallowed
+				continue
+			}
+			conn, dialErr := dialer.DialContext(ctx, network, net.JoinHostPort(ip.String(), port))
+			if dialErr != nil {
+				lastErr = dialErr
+				continue
+			}
+			return conn, nil
+		}
+		return nil, lastErr
+	}
+
+	return &http.Client{
+		Timeout:   providerRequestTimeout,
+		Transport: transport,
+		CheckRedirect: func(*http.Request, []*http.Request) error {
+			return http.ErrUseLastResponse
+		},
+	}
+}
+
+func providerIPAllowed(policy providerEgressPolicy, ip net.IP) bool {
+	if ip == nil || ip.IsUnspecified() || ip.IsLinkLocalUnicast() || ip.IsLinkLocalMulticast() ||
+		ip.IsMulticast() || providerReservedIP(ip) {
+		return false
+	}
+	if ip.IsLoopback() && !policy.allowLoopback {
+		return false
+	}
+	if ip.IsPrivate() && !policy.allowPrivate {
+		return false
+	}
+	return true
+}
+
+// Ranges not classified as private by net.IP but unsuitable for provider
+// egress: CGNAT, benchmarking, and reserved future-use IPv4 space.
+var providerReservedCIDRs = parseProviderCIDRs("100.64.0.0/10", "198.18.0.0/15", "240.0.0.0/4")
+
+func parseProviderCIDRs(specs ...string) []*net.IPNet {
+	networks := make([]*net.IPNet, 0, len(specs))
+	for _, spec := range specs {
+		_, network, err := net.ParseCIDR(spec)
+		if err != nil {
+			panic(err)
+		}
+		networks = append(networks, network)
+	}
+	return networks
+}
+
+func providerReservedIP(ip net.IP) bool {
+	for _, network := range providerReservedCIDRs {
+		if network.Contains(ip) {
+			return true
+		}
+	}
+	return false
+}
diff --git a/server/internal/domain/chat/egress_test.go b/server/internal/domain/chat/egress_test.go
new file mode 100644
index 000000000..44c84a5bc
--- /dev/null
+++ b/server/internal/domain/chat/egress_test.go
@@ -0,0 +1,72 @@
+package chat
+
+import (
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestProviderIPPolicy(t *testing.T) {
+	publicPolicy := providerEgressPolicyFor(ProviderOpenAI, ProviderEgressConfig{})
+	ollamaPolicy := providerEgressPolicyFor(ProviderOllama, ProviderEgressConfig{})
+	privateOllamaPolicy := providerEgressPolicyFor(ProviderOllama, ProviderEgressConfig{AllowPrivateOllama: true})
+
+	assert.False(t, providerIPAllowed(publicPolicy, net.ParseIP("127.0.0.1")))
+	assert.False(t, providerIPAllowed(publicPolicy, net.ParseIP("10.0.0.1")))
+	assert.True(t, providerIPAllowed(publicPolicy, net.ParseIP("203.0.113.10")))
+
+	assert.True(t, providerIPAllowed(ollamaPolicy, net.ParseIP("127.0.0.1")))
+	assert.False(t, providerIPAllowed(ollamaPolicy, net.ParseIP("10.0.0.1")))
+	assert.True(t, providerIPAllowed(privateOllamaPolicy, net.ParseIP("10.0.0.1")))
+	assert.False(t, providerIPAllowed(ollamaPolicy, net.ParseIP("169.254.169.254")))
+	assert.False(t, providerIPAllowed(privateOllamaPolicy, net.ParseIP("169.254.169.254")))
+	assert.False(t, providerIPAllowed(ollamaPolicy, net.ParseIP("198.18.0.1")))
+}
+
+func TestProviderHTTPClientRejectsPublicProviderLoopback(t *testing.T) {
+	client := newProviderHTTPClient(providerEgressPolicyFor(ProviderOpenAI, ProviderEgressConfig{}))
+
+	response, err := client.Get("https://127.0.0.1:4444")
+	if response != nil {
+		response.Body.Close()
+	}
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "disallowed internal address")
+}
+
+func TestProviderHTTPClientRejectsPrivateOllamaByDefault(t *testing.T) {
+	client := newProviderHTTPClient(providerEgressPolicyFor(ProviderOllama, ProviderEgressConfig{}))
+
+	response, err := client.Get("http://10.0.0.1:11434")
+	if response != nil {
+		response.Body.Close()
+	}
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "disallowed internal address")
+}
+
+func TestProviderHTTPClientDoesNotFollowRedirects(t *testing.T) {
+	redirectFollowed := false
+	target := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {
+		redirectFollowed = true
+	}))
+	defer target.Close()
+	redirect := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		http.Redirect(w, req, target.URL, http.StatusFound)
+	}))
+	defer redirect.Close()
+
+	client := newProviderHTTPClient(providerEgressPolicyFor(ProviderOllama, ProviderEgressConfig{}))
+	response, err := client.Get(redirect.URL)
+	require.NoError(t, err)
+	defer response.Body.Close()
+
+	assert.Equal(t, http.StatusFound, response.StatusCode)
+	assert.False(t, redirectFollowed)
+}
diff --git a/server/internal/domain/chat/provider.go b/server/internal/domain/chat/provider.go
new file mode 100644
index 000000000..b7c35893f
--- /dev/null
+++ b/server/internal/domain/chat/provider.go
@@ -0,0 +1,496 @@
+package chat
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/block/proto-fleet/server/internal/domain/fleeterror"
+	"github.com/google/uuid"
+)
+
+const maxProviderResponseBytes = 4 << 20
+
+// OpenAI's model-list response identifies models but does not expose endpoint
+// or function-calling capabilities. Keep discovery conservative: the native
+// agent currently depends on both Chat Completions and function calling.
+var openAIChatAgentModelFamilies = []string{
+	"gpt-5.6-sol",
+	"gpt-5.6-terra",
+	"gpt-5.6-luna",
+	"gpt-5.6",
+	"gpt-5.5",
+	"gpt-5.4-mini",
+	"gpt-5.4-nano",
+	"gpt-5.4",
+	"gpt-5.2",
+	"gpt-5.1",
+	"gpt-5-mini",
+	"gpt-5-nano",
+	"gpt-5",
+	"gpt-4.1-mini",
+	"gpt-4.1",
+	"gpt-4o-mini",
+}
+
+type HTTPModelClient struct {
+	publicHTTPClient *http.Client
+	ollamaHTTPClient *http.Client
+}
+
+type AvailableModel struct {
+	ID          string
+	DisplayName string
+}
+
+type ModelDiscoverer interface {
+	DiscoverModels(ctx context.Context, config RuntimeConfig) ([]AvailableModel, error)
+}
+
+func NewHTTPModelClient(egressConfig ProviderEgressConfig) *HTTPModelClient {
+	return &HTTPModelClient{
+		publicHTTPClient: newProviderHTTPClient(providerEgressPolicyFor(ProviderOpenAI, egressConfig)),
+		ollamaHTTPClient: newProviderHTTPClient(providerEgressPolicyFor(ProviderOllama, egressConfig)),
+	}
+}
+
+func (c *HTTPModelClient) Complete(ctx context.Context, config RuntimeConfig, messages []Message, tools []ToolDefinition) (Completion, error) {
+	switch config.Provider {
+	case ProviderUnspecified:
+		return Completion{}, fleeterror.NewInvalidArgumentError("select an LLM provider")
+	case ProviderOpenAI, ProviderCustom:
+		return c.completeOpenAI(ctx, config, messages, tools)
+	case ProviderAnthropic:
+		return c.completeAnthropic(ctx, config, messages, tools)
+	case ProviderOllama:
+		return c.completeOllama(ctx, config, messages, tools)
+	default:
+		return Completion{}, fleeterror.NewInvalidArgumentErrorf("unsupported LLM provider %q", config.Provider)
+	}
+}
+
+func (c *HTTPModelClient) DiscoverModels(ctx context.Context, config RuntimeConfig) ([]AvailableModel, error) {
+	switch config.Provider {
+	case ProviderOpenAI:
+		return c.discoverOpenAIModels(ctx, config)
+	case ProviderAnthropic:
+		return c.discoverAnthropicModels(ctx, config)
+	case ProviderOllama:
+		return c.discoverOllamaModels(ctx, config)
+	case ProviderCustom:
+		return nil, fleeterror.NewInvalidArgumentError("model discovery is not available for custom providers")
+	case ProviderUnspecified:
+		return nil, fleeterror.NewInvalidArgumentError("select an LLM provider")
+	default:
+		return nil, fleeterror.NewInvalidArgumentErrorf("unsupported LLM provider %q", config.Provider)
+	}
+}
+
+func (c *HTTPModelClient) discoverOpenAIModels(ctx context.Context, config RuntimeConfig) ([]AvailableModel, error) {
+	var response struct {
+		Data []struct {
+			ID string `json:"id"`
+		} `json:"data"`
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/models")
+	headers := map[string]string{"Authorization": "Bearer " + config.APIKey}
+	if err := c.getJSON(ctx, config.Provider, endpoint, headers, &response); err != nil {
+		return nil, err
+	}
+	models := make([]AvailableModel, 0, len(response.Data))
+	for _, model := range response.Data {
+		if !supportsOpenAIChatAgentFlow(model.ID) {
+			continue
+		}
+		models = append(models, AvailableModel{ID: model.ID, DisplayName: model.ID})
+	}
+	return normalizeAvailableModels(models), nil
+}
+
+func supportsOpenAIChatAgentFlow(modelID string) bool {
+	modelID = strings.ToLower(strings.TrimSpace(modelID))
+	for _, family := range openAIChatAgentModelFamilies {
+		if modelID == family || strings.HasPrefix(modelID, family+"-20") {
+			return true
+		}
+	}
+	return false
+}
+
+func (c *HTTPModelClient) discoverAnthropicModels(ctx context.Context, config RuntimeConfig) ([]AvailableModel, error) {
+	var response struct {
+		Data []struct {
+			ID          string `json:"id"`
+			DisplayName string `json:"display_name"`
+		} `json:"data"`
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/v1/models") + "?limit=1000"
+	headers := map[string]string{
+		"x-api-key":         config.APIKey,
+		"anthropic-version": "2023-06-01",
+	}
+	if err := c.getJSON(ctx, config.Provider, endpoint, headers, &response); err != nil {
+		return nil, err
+	}
+	models := make([]AvailableModel, 0, len(response.Data))
+	for _, model := range response.Data {
+		models = append(models, AvailableModel{ID: model.ID, DisplayName: model.DisplayName})
+	}
+	return normalizeAvailableModels(models), nil
+}
+
+func (c *HTTPModelClient) discoverOllamaModels(ctx context.Context, config RuntimeConfig) ([]AvailableModel, error) {
+	var response struct {
+		Models []struct {
+			Name  string `json:"name"`
+			Model string `json:"model"`
+		} `json:"models"`
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/api/tags")
+	if err := c.getJSON(ctx, config.Provider, endpoint, nil, &response); err != nil {
+		return nil, err
+	}
+	models := make([]AvailableModel, 0, len(response.Models))
+	for _, model := range response.Models {
+		id := model.Model
+		if id == "" {
+			id = model.Name
+		}
+		models = append(models, AvailableModel{ID: id, DisplayName: model.Name})
+	}
+	return normalizeAvailableModels(models), nil
+}
+
+func normalizeAvailableModels(models []AvailableModel) []AvailableModel {
+	seen := make(map[string]struct{}, len(models))
+	normalized := make([]AvailableModel, 0, len(models))
+	for _, model := range models {
+		model.ID = strings.TrimSpace(model.ID)
+		model.DisplayName = strings.TrimSpace(model.DisplayName)
+		if model.ID == "" {
+			continue
+		}
+		if _, ok := seen[model.ID]; ok {
+			continue
+		}
+		seen[model.ID] = struct{}{}
+		if model.DisplayName == "" {
+			model.DisplayName = model.ID
+		}
+		normalized = append(normalized, model)
+	}
+	return normalized
+}
+
+type openAIMessage struct {
+	Role       string           `json:"role"`
+	Content    string           `json:"content,omitempty"`
+	ToolCalls  []openAIToolCall `json:"tool_calls,omitempty"`
+	ToolCallID string           `json:"tool_call_id,omitempty"`
+}
+
+type openAIToolCall struct {
+	ID       string             `json:"id"`
+	Type     string             `json:"type"`
+	Function openAIFunctionCall `json:"function"`
+}
+
+type openAIFunctionCall struct {
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+}
+
+func toOpenAIMessages(messages []Message) []openAIMessage {
+	out := make([]openAIMessage, 0, len(messages))
+	for _, message := range messages {
+		converted := openAIMessage{
+			Role:       message.Role,
+			Content:    message.Content,
+			ToolCallID: message.ToolCallID,
+		}
+		for _, call := range message.ToolCalls {
+			converted.ToolCalls = append(converted.ToolCalls, openAIToolCall{
+				ID:   call.ID,
+				Type: "function",
+				Function: openAIFunctionCall{
+					Name:      call.Name,
+					Arguments: string(call.Arguments),
+				},
+			})
+		}
+		out = append(out, converted)
+	}
+	return out
+}
+
+func openAITools(tools []ToolDefinition) []map[string]any {
+	out := make([]map[string]any, 0, len(tools))
+	for _, tool := range tools {
+		out = append(out, map[string]any{
+			"type": "function",
+			"function": map[string]any{
+				"name":        tool.Name,
+				"description": tool.Description,
+				"parameters":  tool.InputSchema,
+			},
+		})
+	}
+	return out
+}
+
+func (c *HTTPModelClient) completeOpenAI(ctx context.Context, config RuntimeConfig, messages []Message, tools []ToolDefinition) (Completion, error) {
+	payload := map[string]any{
+		"model":       config.Model,
+		"messages":    toOpenAIMessages(messages),
+		"tools":       openAITools(tools),
+		"tool_choice": "auto",
+	}
+	// GPT-5-family models reject non-default temperature values. Omitting the
+	// field lets OpenAI apply the model's supported default while preserving the
+	// configured temperature for older OpenAI models.
+	if !strings.HasPrefix(strings.ToLower(strings.TrimSpace(config.Model)), "gpt-5") {
+		payload["temperature"] = config.Temperature
+	}
+	var response struct {
+		Choices []struct {
+			Message openAIMessage `json:"message"`
+		} `json:"choices"`
+	}
+	headers := map[string]string{}
+	if config.APIKey != "" {
+		headers["Authorization"] = "Bearer " + config.APIKey
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/chat/completions")
+	if err := c.doJSON(ctx, config.Provider, endpoint, headers, payload, &response); err != nil {
+		return Completion{}, err
+	}
+	if len(response.Choices) == 0 {
+		return Completion{}, fleeterror.NewUnavailableErrorf("LLM provider returned no choices")
+	}
+	message := response.Choices[0].Message
+	completion := Completion{Content: message.Content}
+	for _, call := range message.ToolCalls {
+		arguments := json.RawMessage(call.Function.Arguments)
+		if !json.Valid(arguments) {
+			arguments = json.RawMessage(`{}`)
+		}
+		id := call.ID
+		if id == "" {
+			id = newToolCallID()
+		}
+		completion.ToolCalls = append(completion.ToolCalls, ModelToolCall{ID: id, Name: call.Function.Name, Arguments: arguments})
+	}
+	return completion, nil
+}
+
+type anthropicBlock struct {
+	Type      string          `json:"type"`
+	Text      string          `json:"text,omitempty"`
+	ID        string          `json:"id,omitempty"`
+	Name      string          `json:"name,omitempty"`
+	Input     json.RawMessage `json:"input,omitempty"`
+	ToolUseID string          `json:"tool_use_id,omitempty"`
+	Content   string          `json:"content,omitempty"`
+	IsError   bool            `json:"is_error,omitempty"`
+}
+
+type anthropicMessage struct {
+	Role    string           `json:"role"`
+	Content []anthropicBlock `json:"content"`
+}
+
+func toAnthropicMessages(messages []Message) (string, []anthropicMessage) {
+	var system string
+	out := make([]anthropicMessage, 0, len(messages))
+	for _, message := range messages {
+		if message.Role == "system" {
+			system = message.Content
+			continue
+		}
+		if message.Role == "tool" {
+			out = append(out, anthropicMessage{Role: "user", Content: []anthropicBlock{{
+				Type:      "tool_result",
+				ToolUseID: message.ToolCallID,
+				Content:   message.Content,
+			}}})
+			continue
+		}
+		converted := anthropicMessage{Role: message.Role}
+		if message.Content != "" {
+			converted.Content = append(converted.Content, anthropicBlock{Type: "text", Text: message.Content})
+		}
+		for _, call := range message.ToolCalls {
+			converted.Content = append(converted.Content, anthropicBlock{Type: "tool_use", ID: call.ID, Name: call.Name, Input: call.Arguments})
+		}
+		out = append(out, converted)
+	}
+	return system, out
+}
+
+func (c *HTTPModelClient) completeAnthropic(ctx context.Context, config RuntimeConfig, messages []Message, tools []ToolDefinition) (Completion, error) {
+	system, converted := toAnthropicMessages(messages)
+	anthropicTools := make([]map[string]any, 0, len(tools))
+	for _, tool := range tools {
+		anthropicTools = append(anthropicTools, map[string]any{
+			"name":         tool.Name,
+			"description":  tool.Description,
+			"input_schema": tool.InputSchema,
+		})
+	}
+	payload := map[string]any{
+		"model":       config.Model,
+		"system":      system,
+		"messages":    converted,
+		"temperature": config.Temperature,
+		"max_tokens":  2048,
+		"tools":       anthropicTools,
+	}
+	var response struct {
+		Content []anthropicBlock `json:"content"`
+	}
+	headers := map[string]string{
+		"x-api-key":         config.APIKey,
+		"anthropic-version": "2023-06-01",
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/v1/messages")
+	if err := c.doJSON(ctx, config.Provider, endpoint, headers, payload, &response); err != nil {
+		return Completion{}, err
+	}
+	completion := Completion{}
+	for _, block := range response.Content {
+		switch block.Type {
+		case "text":
+			completion.Content += block.Text
+		case "tool_use":
+			arguments := block.Input
+			if !json.Valid(arguments) {
+				arguments = json.RawMessage(`{}`)
+			}
+			id := block.ID
+			if id == "" {
+				id = newToolCallID()
+			}
+			completion.ToolCalls = append(completion.ToolCalls, ModelToolCall{ID: id, Name: block.Name, Arguments: arguments})
+		}
+	}
+	return completion, nil
+}
+
+func (c *HTTPModelClient) completeOllama(ctx context.Context, config RuntimeConfig, messages []Message, tools []ToolDefinition) (Completion, error) {
+	payload := map[string]any{
+		"model":    config.Model,
+		"messages": toOpenAIMessages(messages),
+		"stream":   false,
+		"tools":    openAITools(tools),
+		"options":  map[string]any{"temperature": config.Temperature},
+	}
+	var response struct {
+		Message struct {
+			Content   string `json:"content"`
+			ToolCalls []struct {
+				Function struct {
+					Name      string          `json:"name"`
+					Arguments json.RawMessage `json:"arguments"`
+				} `json:"function"`
+			} `json:"tool_calls"`
+		} `json:"message"`
+	}
+	endpoint := providerEndpoint(config.BaseURL, "/api/chat")
+	if err := c.doJSON(ctx, config.Provider, endpoint, nil, payload, &response); err != nil {
+		return Completion{}, err
+	}
+	completion := Completion{Content: response.Message.Content}
+	for _, call := range response.Message.ToolCalls {
+		arguments := call.Function.Arguments
+		if !json.Valid(arguments) {
+			arguments = json.RawMessage(`{}`)
+		}
+		completion.ToolCalls = append(completion.ToolCalls, ModelToolCall{
+			ID:        newToolCallID(),
+			Name:      call.Function.Name,
+			Arguments: arguments,
+		})
+	}
+	return completion, nil
+}
+
+func newToolCallID() string {
+	return "tool-" + uuid.NewString()
+}
+
+func providerEndpoint(baseURL, suffix string) string {
+	base := strings.TrimRight(baseURL, "/")
+	if strings.HasSuffix(base, suffix) {
+		return base
+	}
+	if suffix == "/chat/completions" && strings.HasSuffix(base, "/v1") {
+		return base + suffix
+	}
+	return base + suffix
+}
+
+func (c *HTTPModelClient) doJSON(ctx context.Context, provider Provider, endpoint string, headers map[string]string, payload, target any) error {
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return fleeterror.NewInternalErrorf("marshal LLM request: %v", err)
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, endpoint, bytes.NewReader(body))
+	if err != nil {
+		return fleeterror.NewInternalErrorf("create LLM request: %v", err)
+	}
+	req.Header.Set("Content-Type", "application/json")
+	return c.executeJSON(provider, req, headers, target)
+}
+
+func (c *HTTPModelClient) getJSON(ctx context.Context, provider Provider, endpoint string, headers map[string]string, target any) error {
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, endpoint, nil)
+	if err != nil {
+		return fleeterror.NewInternalErrorf("create LLM request: %v", err)
+	}
+	return c.executeJSON(provider, req, headers, target)
+}
+
+func (c *HTTPModelClient) executeJSON(provider Provider, req *http.Request, headers map[string]string, target any) error {
+	if provider != ProviderOllama && req.URL.Scheme != "https" {
+		return fleeterror.NewUnavailableErrorf("LLM provider request refused: HTTPS is required")
+	}
+	for name, value := range headers {
+		if value != "" {
+			req.Header.Set(name, value)
+		}
+	}
+	httpClient := c.publicHTTPClient
+	if provider == ProviderOllama {
+		httpClient = c.ollamaHTTPClient
+	}
+	resp, err := httpClient.Do(req)
+	if err != nil {
+		if errors.Is(err, errProviderDestinationDisallowed) {
+			return fleeterror.NewUnavailableErrorf("LLM provider destination is blocked by the server egress policy")
+		}
+		return fleeterror.NewUnavailableErrorf("LLM provider request failed")
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		// Provider-controlled response bodies are deliberately left unread. A
+		// provider or intermediary can reflect request headers, including API
+		// keys, in its error payload; returning that text would expose the secret
+		// to the browser and error-logging interceptors.
+		return fleeterror.NewUnavailableErrorf("LLM provider returned HTTP %d", resp.StatusCode)
+	}
+	responseBody, err := io.ReadAll(io.LimitReader(resp.Body, maxProviderResponseBytes+1))
+	if err != nil {
+		return fleeterror.NewUnavailableErrorf("LLM provider response could not be read")
+	}
+	if len(responseBody) > maxProviderResponseBytes {
+		return fleeterror.NewUnavailableErrorf("LLM provider response exceeded %d bytes", maxProviderResponseBytes)
+	}
+	if err := json.Unmarshal(responseBody, target); err != nil {
+		return fleeterror.NewUnavailableErrorf("LLM provider returned an invalid JSON response")
+	}
+	return nil
+}
diff --git a/server/internal/domain/chat/provider_test.go b/server/internal/domain/chat/provider_test.go
new file mode 100644
index 000000000..f7af3be74
--- /dev/null
+++ b/server/internal/domain/chat/provider_test.go
@@ -0,0 +1,229 @@
+package chat
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func newTestHTTPModelClient(client *http.Client) *HTTPModelClient {
+	return &HTTPModelClient{publicHTTPClient: client, ollamaHTTPClient: client}
+}
+
+func TestHTTPModelClientOpenAIRequestAndToolResponse(t *testing.T) {
+	server := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		assert.Equal(t, "/v1/chat/completions", req.URL.Path)
+		assert.Equal(t, "Bearer test-key", req.Header.Get("Authorization"))
+		var body map[string]any
+		require.NoError(t, json.NewDecoder(req.Body).Decode(&body))
+		assert.Equal(t, "test-model", body["model"])
+		assert.Equal(t, 0.3, body["temperature"])
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"choices":[{"message":{"role":"assistant","tool_calls":[{"id":"call-1","type":"function","function":{"name":"list_sites","arguments":"{}"}}]}}]}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	client := newTestHTTPModelClient(server.Client())
+	completion, err := client.Complete(t.Context(), RuntimeConfig{
+		Provider:    ProviderOpenAI,
+		APIKey:      "test-key",
+		BaseURL:     server.URL + "/v1",
+		Model:       "test-model",
+		Temperature: 0.3,
+	}, []Message{{Role: "user", Content: "List sites"}}, []ToolDefinition{{
+		Name: "list_sites", InputSchema: map[string]any{"type": "object"},
+	}})
+
+	require.NoError(t, err)
+	require.Len(t, completion.ToolCalls, 1)
+	assert.Equal(t, "call-1", completion.ToolCalls[0].ID)
+	assert.Equal(t, "list_sites", completion.ToolCalls[0].Name)
+}
+
+func TestHTTPModelClientOmitsTemperatureForGPT5(t *testing.T) {
+	server := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		var body map[string]any
+		require.NoError(t, json.NewDecoder(req.Body).Decode(&body))
+		assert.Equal(t, "gpt-5", body["model"])
+		assert.NotContains(t, body, "temperature")
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"choices":[{"message":{"role":"assistant","content":"Hello"}}]}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	client := newTestHTTPModelClient(server.Client())
+	completion, err := client.Complete(t.Context(), RuntimeConfig{
+		Provider:    ProviderOpenAI,
+		BaseURL:     server.URL + "/v1",
+		Model:       "gpt-5",
+		Temperature: 0.2,
+	}, []Message{{Role: "user", Content: "Hello"}}, nil)
+
+	require.NoError(t, err)
+	assert.Equal(t, "Hello", completion.Content)
+}
+
+func TestHTTPModelClientDoesNotEchoProviderErrorBodiesOrReflectedCredentials(t *testing.T) {
+	server := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusBadRequest)
+		_, err := w.Write([]byte(`{"error":{"message":"rejected credential secret-key"}}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	client := newTestHTTPModelClient(server.Client())
+	_, err := client.Complete(t.Context(), RuntimeConfig{
+		Provider: ProviderOpenAI,
+		APIKey:   "secret-key",
+		BaseURL:  server.URL,
+		Model:    "test-model",
+	}, []Message{{Role: "user", Content: "Hello"}}, nil)
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "HTTP 400")
+	assert.NotContains(t, err.Error(), "rejected credential")
+	assert.NotContains(t, err.Error(), "secret-key")
+}
+
+func TestHTTPModelClientDiscoversOpenAIModels(t *testing.T) {
+	server := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		assert.Equal(t, http.MethodGet, req.Method)
+		assert.Equal(t, "/v1/models", req.URL.Path)
+		assert.Equal(t, "Bearer test-key", req.Header.Get("Authorization"))
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"data":[{"id":"gpt-5.6"},{"id":"gpt-5.6"},{"id":"gpt-5.5-pro"},{"id":"gpt-5.4-2026-03-05"},{"id":"gpt-5.4-pro"},{"id":"gpt-5.3-codex"},{"id":"gpt-image-2"},{"id":"text-embedding-3-small"},{"id":""}]}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	models, err := newTestHTTPModelClient(server.Client()).DiscoverModels(t.Context(), RuntimeConfig{
+		Provider: ProviderOpenAI,
+		APIKey:   "test-key",
+		BaseURL:  server.URL + "/v1",
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, []AvailableModel{
+		{ID: "gpt-5.6", DisplayName: "gpt-5.6"},
+		{ID: "gpt-5.4-2026-03-05", DisplayName: "gpt-5.4-2026-03-05"},
+	}, models)
+}
+
+func TestSupportsOpenAIChatAgentFlow(t *testing.T) {
+	testCases := []struct {
+		name      string
+		modelID   string
+		supported bool
+	}{
+		{name: "current alias", modelID: "gpt-5.6", supported: true},
+		{name: "current tier", modelID: "gpt-5.6-terra", supported: true},
+		{name: "dated snapshot", modelID: "gpt-5.4-mini-2026-03-17", supported: true},
+		{name: "older agent model", modelID: "gpt-4.1-mini", supported: true},
+		{name: "responses-only pro", modelID: "gpt-5.5-pro", supported: false},
+		{name: "responses-only pro snapshot", modelID: "gpt-5.4-pro-2026-03-05", supported: false},
+		{name: "responses-only codex", modelID: "gpt-5.3-codex", supported: false},
+		{name: "specialized audio", modelID: "gpt-4o-mini-transcribe", supported: false},
+		{name: "image generation", modelID: "gpt-image-2", supported: false},
+		{name: "embedding", modelID: "text-embedding-3-small", supported: false},
+	}
+
+	for _, testCase := range testCases {
+		t.Run(testCase.name, func(t *testing.T) {
+			assert.Equal(t, testCase.supported, supportsOpenAIChatAgentFlow(testCase.modelID))
+		})
+	}
+}
+
+func TestHTTPModelClientDiscoversAnthropicModels(t *testing.T) {
+	server := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		assert.Equal(t, http.MethodGet, req.Method)
+		assert.Equal(t, "/v1/models", req.URL.Path)
+		assert.Equal(t, "1000", req.URL.Query().Get("limit"))
+		assert.Equal(t, "test-key", req.Header.Get("X-Api-Key"))
+		assert.Equal(t, "2023-06-01", req.Header.Get("Anthropic-Version"))
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"data":[{"id":"claude-test","display_name":"Claude Test"}]}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	models, err := newTestHTTPModelClient(server.Client()).DiscoverModels(t.Context(), RuntimeConfig{
+		Provider: ProviderAnthropic,
+		APIKey:   "test-key",
+		BaseURL:  server.URL,
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, []AvailableModel{{ID: "claude-test", DisplayName: "Claude Test"}}, models)
+}
+
+func TestHTTPModelClientDiscoversInstalledOllamaModels(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		assert.Equal(t, http.MethodGet, req.Method)
+		assert.Equal(t, "/api/tags", req.URL.Path)
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"models":[{"name":"llama-test:latest","model":"llama-test:latest"}]}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	models, err := NewHTTPModelClient(ProviderEgressConfig{}).DiscoverModels(t.Context(), RuntimeConfig{
+		Provider: ProviderOllama,
+		BaseURL:  server.URL,
+	})
+
+	require.NoError(t, err)
+	assert.Equal(t, []AvailableModel{{ID: "llama-test:latest", DisplayName: "llama-test:latest"}}, models)
+}
+
+func TestHTTPModelClientReportsServerPolicyForPrivateOllamaDestination(t *testing.T) {
+	_, err := NewHTTPModelClient(ProviderEgressConfig{}).DiscoverModels(t.Context(), RuntimeConfig{
+		Provider: ProviderOllama,
+		BaseURL:  "http://10.0.0.20:11434",
+	})
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "blocked by the server egress policy")
+	assert.NotContains(t, err.Error(), "10.0.0.20")
+}
+
+func TestHTTPModelClientGeneratesUniqueOllamaToolCallIDs(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, err := w.Write([]byte(`{"message":{"tool_calls":[{"function":{"name":"list_sites","arguments":{}}}]}}`))
+		require.NoError(t, err)
+	}))
+	defer server.Close()
+
+	client := NewHTTPModelClient(ProviderEgressConfig{})
+	config := RuntimeConfig{Provider: ProviderOllama, BaseURL: server.URL, Model: "llama-test"}
+	first, err := client.Complete(t.Context(), config, []Message{{Role: "user", Content: "List sites"}}, nil)
+	require.NoError(t, err)
+	second, err := client.Complete(t.Context(), config, []Message{{Role: "user", Content: "List sites again"}}, nil)
+	require.NoError(t, err)
+
+	require.Len(t, first.ToolCalls, 1)
+	require.Len(t, second.ToolCalls, 1)
+	assert.NotEmpty(t, first.ToolCalls[0].ID)
+	assert.NotEqual(t, first.ToolCalls[0].ID, second.ToolCalls[0].ID)
+}
+
+func TestHTTPModelClientRefusesPlainHTTPForCredentialProviders(t *testing.T) {
+	client := NewHTTPModelClient(ProviderEgressConfig{})
+
+	_, err := client.Complete(t.Context(), RuntimeConfig{
+		Provider: ProviderOpenAI,
+		BaseURL:  "http://203.0.113.10/v1",
+		Model:    "gpt-test",
+	}, []Message{{Role: "user", Content: "Hello"}}, nil)
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), "HTTPS is required")
+}
diff --git a/server/internal/domain/stores/sqlstores/llm_config.go b/server/internal/domain/stores/sqlstores/llm_config.go
new file mode 100644
index 000000000..289885254
--- /dev/null
+++ b/server/internal/domain/stores/sqlstores/llm_config.go
@@ -0,0 +1,59 @@
+package sqlstores
+
+import (
+	"context"
+	"database/sql"
+
+	"github.com/block/proto-fleet/server/generated/sqlc"
+	"github.com/block/proto-fleet/server/internal/domain/chat"
+)
+
+type SQLLLMConfigStore struct {
+	SQLConnectionManager
+}
+
+func NewSQLLLMConfigStore(conn *sql.DB) *SQLLLMConfigStore {
+	return &SQLLLMConfigStore{SQLConnectionManager: NewSQLConnectionManager(conn)}
+}
+
+var _ chat.ConfigStore = (*SQLLLMConfigStore)(nil)
+
+func llmConfigToRecord(row sqlc.LlmConfig) chat.ConfigRecord {
+	return chat.ConfigRecord{
+		OrganizationID:       row.OrganizationID,
+		Harness:              chat.Harness(row.Harness),
+		Provider:             chat.Provider(row.Provider),
+		APIKeyEncrypted:      row.ApiKeyEncrypted,
+		BaseURL:              row.BaseUrl,
+		Model:                row.Model,
+		Temperature:          row.Temperature,
+		GooseBaseURL:         row.GooseBaseUrl,
+		GooseSecretEncrypted: row.GooseSecretEncrypted,
+	}
+}
+
+func (s *SQLLLMConfigStore) Get(ctx context.Context, orgID int64) (chat.ConfigRecord, error) {
+	row, err := s.GetQueries(ctx).GetLLMConfig(ctx, orgID)
+	if err != nil {
+		return chat.ConfigRecord{}, err
+	}
+	return llmConfigToRecord(row), nil
+}
+
+func (s *SQLLLMConfigStore) Upsert(ctx context.Context, record chat.ConfigRecord) (chat.ConfigRecord, error) {
+	row, err := s.GetQueries(ctx).UpsertLLMConfig(ctx, sqlc.UpsertLLMConfigParams{
+		OrganizationID:       record.OrganizationID,
+		Harness:              string(record.Harness),
+		Provider:             string(record.Provider),
+		ApiKeyEncrypted:      record.APIKeyEncrypted,
+		BaseUrl:              record.BaseURL,
+		Model:                record.Model,
+		Temperature:          record.Temperature,
+		GooseBaseUrl:         record.GooseBaseURL,
+		GooseSecretEncrypted: record.GooseSecretEncrypted,
+	})
+	if err != nil {
+		return chat.ConfigRecord{}, err
+	}
+	return llmConfigToRecord(row), nil
+}
diff --git a/server/internal/handlers/chat/handler.go b/server/internal/handlers/chat/handler.go
new file mode 100644
index 000000000..25a33b10f
--- /dev/null
+++ b/server/internal/handlers/chat/handler.go
@@ -0,0 +1,225 @@
+package chat
+
+import (
+	"context"
+
+	"connectrpc.com/connect"
+	"github.com/google/uuid"
+
+	chatv1 "github.com/block/proto-fleet/server/generated/grpc/chat/v1"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
+	"github.com/block/proto-fleet/server/internal/domain/authz"
+	chatdomain "github.com/block/proto-fleet/server/internal/domain/chat"
+	"github.com/block/proto-fleet/server/internal/handlers/middleware"
+)
+
+type Handler struct {
+	config        *chatdomain.ConfigService
+	agent         *chatdomain.Agent
+	discoverer    chatdomain.ModelDiscoverer
+	tools         chatdomain.ToolRegistry
+	confirmations *chatdomain.ConfirmationBroker
+}
+
+func NewHandler(
+	config *chatdomain.ConfigService,
+	agent *chatdomain.Agent,
+	discoverer chatdomain.ModelDiscoverer,
+	tools chatdomain.ToolRegistry,
+	confirmations *chatdomain.ConfirmationBroker,
+) *Handler {
+	return &Handler{config: config, agent: agent, discoverer: discoverer, tools: tools, confirmations: confirmations}
+}
+
+var _ chatv1connect.ChatServiceHandler = (*Handler)(nil)
+
+func (h *Handler) GetLLMConfig(ctx context.Context, _ *connect.Request[chatv1.GetLLMConfigRequest]) (*connect.Response[chatv1.GetLLMConfigResponse], error) {
+	info, err := middleware.RequirePermission(ctx, authz.PermAPIKeyManage, authz.ResourceContext{})
+	if err != nil {
+		return nil, err
+	}
+	config, err := h.config.Get(ctx, info.OrganizationID)
+	if err != nil {
+		return nil, err
+	}
+	return connect.NewResponse(&chatv1.GetLLMConfigResponse{Config: configToProto(config)}), nil
+}
+
+func (h *Handler) DiscoverModels(ctx context.Context, req *connect.Request[chatv1.DiscoverModelsRequest]) (*connect.Response[chatv1.DiscoverModelsResponse], error) {
+	info, err := middleware.RequirePermission(ctx, authz.PermAPIKeyManage, authz.ResourceContext{})
+	if err != nil {
+		return nil, err
+	}
+	config, err := h.config.DiscoveryRuntime(ctx, info.OrganizationID, chatdomain.DiscoverConfig{
+		Provider:        providerFromProto(req.Msg.GetProvider()),
+		APIKey:          req.Msg.GetApiKey(),
+		BaseURL:         req.Msg.GetBaseUrl(),
+		UseStoredAPIKey: req.Msg.GetUseStoredApiKey(),
+	})
+	if err != nil {
+		return nil, err
+	}
+	models, err := h.discoverer.DiscoverModels(ctx, config)
+	if err != nil {
+		return nil, err
+	}
+	response := &chatv1.DiscoverModelsResponse{Models: make([]*chatv1.AvailableModel, 0, len(models))}
+	for _, model := range models {
+		response.Models = append(response.Models, &chatv1.AvailableModel{Id: model.ID, DisplayName: model.DisplayName})
+	}
+	return connect.NewResponse(response), nil
+}
+
+func (h *Handler) UpdateLLMConfig(ctx context.Context, req *connect.Request[chatv1.UpdateLLMConfigRequest]) (*connect.Response[chatv1.UpdateLLMConfigResponse], error) {
+	info, err := middleware.RequirePermission(ctx, authz.PermAPIKeyManage, authz.ResourceContext{})
+	if err != nil {
+		return nil, err
+	}
+	config, err := h.config.Update(ctx, info.OrganizationID, chatdomain.UpdateConfig{
+		Harness:          harnessFromProto(req.Msg.GetHarness()),
+		Provider:         providerFromProto(req.Msg.GetProvider()),
+		APIKey:           req.Msg.GetApiKey(),
+		BaseURL:          req.Msg.GetBaseUrl(),
+		Model:            req.Msg.GetModel(),
+		GooseBaseURL:     req.Msg.GetGooseBaseUrl(),
+		GooseSecret:      req.Msg.GetGooseSecret(),
+		ClearAPIKey:      req.Msg.GetClearApiKey(),
+		ClearGooseSecret: req.Msg.GetClearGooseSecret(),
+	})
+	if err != nil {
+		return nil, err
+	}
+	return connect.NewResponse(&chatv1.UpdateLLMConfigResponse{Config: configToProto(config)}), nil
+}
+
+func (h *Handler) SendMessage(ctx context.Context, req *connect.Request[chatv1.SendMessageRequest], stream *connect.ServerStream[chatv1.SendMessageResponse]) error {
+	info, err := middleware.RequirePermission(ctx, authz.PermFleetRead, authz.ResourceContext{})
+	if err != nil {
+		return err
+	}
+	config, err := h.config.Runtime(ctx, info.OrganizationID)
+	if err != nil {
+		return err
+	}
+	history := make([]chatdomain.Message, 0, len(req.Msg.GetHistory()))
+	for _, turn := range req.Msg.GetHistory() {
+		role := "user"
+		if turn.GetRole() == chatv1.ChatRole_CHAT_ROLE_ASSISTANT {
+			role = "assistant"
+		}
+		history = append(history, chatdomain.Message{Role: role, Content: turn.GetContent()})
+	}
+	messageID := uuid.NewString()
+	return h.agent.Run(ctx, config, history, req.Msg.GetContent(), h.tools, func(event chatdomain.Event) error {
+		return stream.Send(eventToProto(messageID, event))
+	})
+}
+
+func (h *Handler) ResolveToolConfirmation(ctx context.Context, req *connect.Request[chatv1.ResolveToolConfirmationRequest]) (*connect.Response[chatv1.ResolveToolConfirmationResponse], error) {
+	if _, err := middleware.RequirePermission(ctx, authz.PermFleetRead, authz.ResourceContext{}); err != nil {
+		return nil, err
+	}
+	decision := chatdomain.ConfirmationCancelled
+	if req.Msg.GetDecision() == chatv1.ToolConfirmationDecision_TOOL_CONFIRMATION_DECISION_APPROVE {
+		decision = chatdomain.ConfirmationApproved
+	}
+	if err := h.confirmations.Resolve(ctx, req.Msg.GetConfirmationId(), decision); err != nil {
+		return nil, err
+	}
+	return connect.NewResponse(&chatv1.ResolveToolConfirmationResponse{}), nil
+}
+
+func configToProto(config chatdomain.ConfigView) *chatv1.LLMConfig {
+	return &chatv1.LLMConfig{
+		Harness:        harnessToProto(config.Harness),
+		Provider:       providerToProto(config.Provider),
+		HasApiKey:      config.HasAPIKey,
+		BaseUrl:        config.BaseURL,
+		Model:          config.Model,
+		GooseBaseUrl:   config.GooseBaseURL,
+		HasGooseSecret: config.HasGooseSecret,
+		Configured:     config.Configured,
+	}
+}
+
+func harnessFromProto(harness chatv1.AgentHarness) chatdomain.Harness {
+	if harness == chatv1.AgentHarness_AGENT_HARNESS_GOOSE {
+		return chatdomain.HarnessGoose
+	}
+	return chatdomain.HarnessNative
+}
+
+func harnessToProto(harness chatdomain.Harness) chatv1.AgentHarness {
+	if harness == chatdomain.HarnessGoose {
+		return chatv1.AgentHarness_AGENT_HARNESS_GOOSE
+	}
+	return chatv1.AgentHarness_AGENT_HARNESS_NATIVE
+}
+
+func providerFromProto(provider chatv1.LLMProvider) chatdomain.Provider {
+	switch provider {
+	case chatv1.LLMProvider_LLM_PROVIDER_UNSPECIFIED:
+		return chatdomain.ProviderUnspecified
+	case chatv1.LLMProvider_LLM_PROVIDER_OPENAI:
+		return chatdomain.ProviderOpenAI
+	case chatv1.LLMProvider_LLM_PROVIDER_ANTHROPIC:
+		return chatdomain.ProviderAnthropic
+	case chatv1.LLMProvider_LLM_PROVIDER_OLLAMA:
+		return chatdomain.ProviderOllama
+	case chatv1.LLMProvider_LLM_PROVIDER_CUSTOM:
+		return chatdomain.ProviderCustom
+	default:
+		return chatdomain.ProviderUnspecified
+	}
+}
+
+func providerToProto(provider chatdomain.Provider) chatv1.LLMProvider {
+	switch provider {
+	case chatdomain.ProviderUnspecified:
+		return chatv1.LLMProvider_LLM_PROVIDER_UNSPECIFIED
+	case chatdomain.ProviderOpenAI:
+		return chatv1.LLMProvider_LLM_PROVIDER_OPENAI
+	case chatdomain.ProviderAnthropic:
+		return chatv1.LLMProvider_LLM_PROVIDER_ANTHROPIC
+	case chatdomain.ProviderOllama:
+		return chatv1.LLMProvider_LLM_PROVIDER_OLLAMA
+	case chatdomain.ProviderCustom:
+		return chatv1.LLMProvider_LLM_PROVIDER_CUSTOM
+	default:
+		return chatv1.LLMProvider_LLM_PROVIDER_UNSPECIFIED
+	}
+}
+
+func eventToProto(messageID string, event chatdomain.Event) *chatv1.SendMessageResponse {
+	response := &chatv1.SendMessageResponse{MessageId: messageID}
+	switch event.Kind {
+	case chatdomain.EventTextDelta:
+		response.Event = &chatv1.SendMessageResponse_TextDelta{TextDelta: &chatv1.TextDelta{Content: event.Content}}
+	case chatdomain.EventToolCall:
+		response.Event = &chatv1.SendMessageResponse_ToolCall{ToolCall: &chatv1.ToolCall{
+			Id: event.ToolCallID, Name: event.ToolName, Summary: event.Summary,
+		}}
+	case chatdomain.EventToolResult:
+		response.Event = &chatv1.SendMessageResponse_ToolResult{ToolResult: &chatv1.ToolResult{
+			Id: event.ToolCallID, Name: event.ToolName, Success: event.Success, Summary: event.Summary, Cancelled: event.Cancelled,
+		}}
+	case chatdomain.EventConfirmationRequired:
+		confirmation := event.Confirmation
+		details := make([]*chatv1.ToolConfirmationDetail, 0, len(confirmation.Details))
+		for _, detail := range confirmation.Details {
+			details = append(details, &chatv1.ToolConfirmationDetail{Label: detail.Label, Value: detail.Value})
+		}
+		response.Event = &chatv1.SendMessageResponse_ConfirmationRequired{ConfirmationRequired: &chatv1.ToolConfirmationRequired{
+			ConfirmationId: event.ConfirmationID,
+			ToolCallId:     event.ToolCallID,
+			ToolName:       event.ToolName,
+			Title:          confirmation.Title,
+			Description:    confirmation.Description,
+			ConfirmLabel:   confirmation.ConfirmLabel,
+			Details:        details,
+		}}
+	case chatdomain.EventDone:
+		response.Event = &chatv1.SendMessageResponse_Done{Done: &chatv1.Done{FinishReason: event.Summary}}
+	}
+	return response
+}
diff --git a/server/internal/handlers/chat/tools.go b/server/internal/handlers/chat/tools.go
new file mode 100644
index 000000000..dd6a1802f
--- /dev/null
+++ b/server/internal/handlers/chat/tools.go
@@ -0,0 +1,2642 @@
+package chat
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"strings"
+
+	"buf.build/go/protovalidate"
+	"connectrpc.com/connect"
+
+	commonv1 "github.com/block/proto-fleet/server/generated/grpc/common/v1"
+	devicesetv1 "github.com/block/proto-fleet/server/generated/grpc/device_set/v1"
+	fleetv1 "github.com/block/proto-fleet/server/generated/grpc/fleetmanagement/v1"
+	minercommandv1 "github.com/block/proto-fleet/server/generated/grpc/minercommand/v1"
+	poolsv1 "github.com/block/proto-fleet/server/generated/grpc/pools/v1"
+	schedulev1 "github.com/block/proto-fleet/server/generated/grpc/schedule/v1"
+	sitesv1 "github.com/block/proto-fleet/server/generated/grpc/sites/v1"
+	chatdomain "github.com/block/proto-fleet/server/internal/domain/chat"
+)
+
+const (
+	defaultResolveMinersLimit = 100
+	maxResolveMinersLimit     = 1000
+	maxResolveMinersScan      = 5000
+	maxRackSlotAssignments    = 1000
+	maxMinerActionDevices     = 1000
+	maxDowntimeTargets        = 1000
+)
+
+const (
+	rackOrderIndexBottomLeft  = "bottom_left"
+	rackOrderIndexTopLeft     = "top_left"
+	rackOrderIndexBottomRight = "bottom_right"
+	rackOrderIndexTopRight    = "top_right"
+	labelUnspecified          = "unspecified"
+)
+
+const (
+	minerActionSelectorExplicit   = "explicit"
+	minerActionSelectorAllDevices = "all_devices"
+	minerActionSelectorFilter     = "filter"
+)
+
+type fleetHandler interface {
+	GetMinerStateCounts(ctx context.Context, req *connect.Request[fleetv1.GetMinerStateCountsRequest]) (*connect.Response[fleetv1.GetMinerStateCountsResponse], error)
+	ListMinerStateSnapshots(ctx context.Context, req *connect.Request[fleetv1.ListMinerStateSnapshotsRequest]) (*connect.Response[fleetv1.ListMinerStateSnapshotsResponse], error)
+}
+
+type sitesHandler interface {
+	ListSites(ctx context.Context, req *connect.Request[sitesv1.ListSitesRequest]) (*connect.Response[sitesv1.ListSitesResponse], error)
+	CreateSite(ctx context.Context, req *connect.Request[sitesv1.CreateSiteRequest]) (*connect.Response[sitesv1.CreateSiteResponse], error)
+}
+
+type poolsHandler interface {
+	ListPools(ctx context.Context, req *connect.Request[poolsv1.ListPoolsRequest]) (*connect.Response[poolsv1.ListPoolsResponse], error)
+}
+
+type deviceSetsHandler interface {
+	GetDeviceSet(ctx context.Context, req *connect.Request[devicesetv1.GetDeviceSetRequest]) (*connect.Response[devicesetv1.GetDeviceSetResponse], error)
+	ListDeviceSets(ctx context.Context, req *connect.Request[devicesetv1.ListDeviceSetsRequest]) (*connect.Response[devicesetv1.ListDeviceSetsResponse], error)
+	ListDeviceSetMembers(ctx context.Context, req *connect.Request[devicesetv1.ListDeviceSetMembersRequest]) (*connect.Response[devicesetv1.ListDeviceSetMembersResponse], error)
+	CreateDeviceSet(ctx context.Context, req *connect.Request[devicesetv1.CreateDeviceSetRequest]) (*connect.Response[devicesetv1.CreateDeviceSetResponse], error)
+	AssignDevicesToRack(ctx context.Context, req *connect.Request[devicesetv1.AssignDevicesToRackRequest]) (*connect.Response[devicesetv1.AssignDevicesToRackResponse], error)
+	SetRackSlotPosition(ctx context.Context, req *connect.Request[devicesetv1.SetRackSlotPositionRequest]) (*connect.Response[devicesetv1.SetRackSlotPositionResponse], error)
+	ClearRackSlotPosition(ctx context.Context, req *connect.Request[devicesetv1.ClearRackSlotPositionRequest]) (*connect.Response[devicesetv1.ClearRackSlotPositionResponse], error)
+	GetRackSlots(ctx context.Context, req *connect.Request[devicesetv1.GetRackSlotsRequest]) (*connect.Response[devicesetv1.GetRackSlotsResponse], error)
+}
+
+type commandHandler interface {
+	CheckCommandCapabilities(ctx context.Context, req *connect.Request[minercommandv1.CheckCommandCapabilitiesRequest]) (*connect.Response[minercommandv1.CheckCommandCapabilitiesResponse], error)
+	Reboot(ctx context.Context, req *connect.Request[minercommandv1.RebootRequest]) (*connect.Response[minercommandv1.RebootResponse], error)
+	StopMining(ctx context.Context, req *connect.Request[minercommandv1.StopMiningRequest]) (*connect.Response[minercommandv1.StopMiningResponse], error)
+	StartMining(ctx context.Context, req *connect.Request[minercommandv1.StartMiningRequest]) (*connect.Response[minercommandv1.StartMiningResponse], error)
+	BlinkLED(ctx context.Context, req *connect.Request[minercommandv1.BlinkLEDRequest]) (*connect.Response[minercommandv1.BlinkLEDResponse], error)
+}
+
+type scheduleHandler interface {
+	CreateSchedule(ctx context.Context, req *connect.Request[schedulev1.CreateScheduleRequest]) (*connect.Response[schedulev1.CreateScheduleResponse], error)
+}
+
+type FleetTools struct {
+	fleet      fleetHandler
+	sites      sitesHandler
+	pools      poolsHandler
+	deviceSets deviceSetsHandler
+	commands   commandHandler
+	schedules  scheduleHandler
+}
+
+func NewFleetTools(
+	fleet fleetHandler,
+	sites sitesHandler,
+	pools poolsHandler,
+	deviceSets deviceSetsHandler,
+	commands commandHandler,
+	schedules scheduleHandler,
+) *FleetTools {
+	return &FleetTools{
+		fleet:      fleet,
+		sites:      sites,
+		pools:      pools,
+		deviceSets: deviceSets,
+		commands:   commands,
+		schedules:  schedules,
+	}
+}
+
+func (t *FleetTools) Definitions() []chatdomain.ToolDefinition {
+	return []chatdomain.ToolDefinition{
+		{
+			Name:        "get_miner_state_counts",
+			Description: "Get current counts of hashing, broken, offline, and sleeping miners, optionally filtered by site IDs.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"properties": map[string]any{
+					"site_ids": map[string]any{
+						"type": "array", "items": map[string]any{"type": "integer"},
+					},
+					"include_unassigned": map[string]any{"type": "boolean"},
+				},
+			},
+		},
+		{
+			Name:        "list_sites",
+			Description: "List sites in the organization with device, building, rack, and infrastructure counts.",
+			InputSchema: emptyObjectSchema(),
+		},
+		{
+			Name:        "get_site_health_summary",
+			Description: "Get a structured health summary for the fleet or one site, including current miner state counts and site inventory rows.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"properties": map[string]any{
+					"site_id":            map[string]any{"type": "integer", "minimum": 1},
+					"include_unassigned": map[string]any{"type": "boolean"},
+				},
+			},
+		},
+		{
+			Name:        "resolve_miners",
+			Description: "Resolve miner descriptions into explicit device identifiers for follow-up write tools. Use this before move_miners_to_rack unless the operator supplied exact device identifiers.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"properties": map[string]any{
+					"query": map[string]any{
+						"type":        "string",
+						"maxLength":   255,
+						"description": "Case-insensitive text matched against returned miner identifiers, names, MACs, serials, IPs, and models after server filters.",
+					},
+					"device_statuses": map[string]any{
+						"type": "array",
+						"items": map[string]any{
+							"type": "string",
+							"enum": []string{
+								"online", "offline", "maintenance", "error", "inactive",
+								"needs_mining_pool", "updating", "reboot_required",
+								"hashing", "broken", "sleeping",
+							},
+						},
+						"uniqueItems": true,
+					},
+					"site_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_unassigned":  map[string]any{"type": "boolean"},
+					"building_ids":        map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_no_building": map[string]any{"type": "boolean"},
+					"rack_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_no_rack":     map[string]any{"type": "boolean"},
+					"models":              map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255}},
+					"ip_cidrs":            map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 64}},
+					"limit":               map[string]any{"type": "integer", "minimum": 1, "maximum": maxResolveMinersLimit, "description": "Maximum miners to return. Use 1000 when resolving all matching miners for a write."},
+				},
+			},
+		},
+		{
+			Name:        "list_actionable_miner_issues",
+			Description: "List miners currently in actionable states with concise suggested next actions. Use this to triage offline, broken, pool-missing, reboot-required, updating, inactive, or maintenance miners.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"properties": map[string]any{
+					"query": map[string]any{
+						"type":        "string",
+						"maxLength":   255,
+						"description": "Optional case-insensitive text matched against returned miner identifiers, names, MACs, serials, IPs, and models after server filters.",
+					},
+					"device_statuses": map[string]any{
+						"type": "array",
+						"items": map[string]any{
+							"type": "string",
+							"enum": []string{
+								"offline", "maintenance", "error", "inactive",
+								"needs_mining_pool", "updating", "reboot_required",
+								"broken", "sleeping",
+							},
+						},
+						"uniqueItems": true,
+						"description": "Defaults to offline, broken, needs_mining_pool, and reboot_required.",
+					},
+					"site_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_unassigned":  map[string]any{"type": "boolean"},
+					"building_ids":        map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_no_building": map[string]any{"type": "boolean"},
+					"rack_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+					"include_no_rack":     map[string]any{"type": "boolean"},
+					"models":              map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255}},
+					"ip_cidrs":            map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 64}},
+					"limit":               map[string]any{"type": "integer", "minimum": 1, "maximum": maxResolveMinersLimit},
+				},
+			},
+		},
+		{
+			Name:        "list_pools",
+			Description: "List saved mining pool names. Connection URLs, usernames, wallet identifiers, worker identifiers, and credentials are never returned.",
+			InputSchema: emptyObjectSchema(),
+		},
+		{
+			Name:        "list_racks",
+			Description: "List racks with IDs, labels, layouts, numbering origin, placement labels, and miner counts so an operator request can be resolved to an exact rack.",
+			InputSchema: emptyObjectSchema(),
+		},
+		{
+			Name:        "get_rack_slots",
+			Description: "List occupied slot positions for one rack. Slot row and column coordinates are 0-indexed.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"rack_id"},
+				"properties": map[string]any{
+					"rack_id": map[string]any{"type": "integer", "minimum": 1},
+				},
+			},
+		},
+		{
+			Name:        "get_rack_health",
+			Description: "Get rack layout, slot occupancy, member list, and current miner status counts for one rack.",
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"rack_id"},
+				"properties": map[string]any{
+					"rack_id": map[string]any{"type": "integer", "minimum": 1},
+					"limit":   map[string]any{"type": "integer", "minimum": 1, "maximum": maxResolveMinersLimit, "description": "Maximum rack member rows to return."},
+				},
+			},
+		},
+		{
+			Name:        "preview_miner_action",
+			Description: "Check whether selected miners support a command before requesting execution. Use selector.type all_devices for whole-fleet actions, filter for backend-resolvable subsets, or explicit for exact miner IDs.",
+			InputSchema: minerActionSchema(),
+		},
+		{
+			Name:                 "execute_miner_action",
+			Description:          "Execute a miner command on selected miners. Use selector.type all_devices for whole-fleet actions, filter for backend-resolvable subsets, or explicit for exact miner IDs. Use preview_miner_action first when possible. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema:          minerActionSchema(),
+		},
+		{
+			Name:        "preview_downtime_window",
+			Description: "Preview a one-time maintenance schedule using the existing schedule service. The sleep action stops mining at the start time and does not automatically resume at end_time.",
+			InputSchema: downtimeWindowSchema(),
+		},
+		{
+			Name:                 "create_downtime_window",
+			Description:          "Create a one-time maintenance schedule using the existing schedule service. The sleep action stops mining at the start time and does not automatically resume at end_time. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema:          downtimeWindowSchema(),
+		},
+		{
+			Name:                 "create_site",
+			Description:          "Create a site. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"name"},
+				"properties": map[string]any{
+					"name":              map[string]any{"type": "string", "minLength": 1, "maxLength": 255},
+					"location_city":     map[string]any{"type": "string", "maxLength": 255},
+					"location_state":    map[string]any{"type": "string", "maxLength": 255},
+					"country":           map[string]any{"type": "string", "maxLength": 2},
+					"address":           map[string]any{"type": "string", "maxLength": 255},
+					"postal_code":       map[string]any{"type": "string", "maxLength": 32},
+					"timezone":          map[string]any{"type": "string", "maxLength": 64},
+					"power_capacity_mw": map[string]any{"type": "number", "minimum": 0},
+					"notes":             map[string]any{"type": "string", "maxLength": 4096},
+				},
+			},
+		},
+		{
+			Name:                 "create_rack",
+			Description:          "Create an empty rack with a grid layout and optional site or building placement. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"label", "rows", "columns"},
+				"properties": map[string]any{
+					"label":        map[string]any{"type": "string", "minLength": 1, "maxLength": 100},
+					"rows":         map[string]any{"type": "integer", "minimum": 1},
+					"columns":      map[string]any{"type": "integer", "minimum": 1},
+					"zone":         map[string]any{"type": "string", "maxLength": 100},
+					"site_id":      map[string]any{"type": "integer", "minimum": 1},
+					"building_id":  map[string]any{"type": "integer", "minimum": 1},
+					"order_index":  map[string]any{"type": "string", "enum": []string{rackOrderIndexTopLeft, rackOrderIndexTopRight, rackOrderIndexBottomLeft, rackOrderIndexBottomRight}},
+					"cooling_type": map[string]any{"type": "string", "enum": []string{"air", "immersion"}},
+				},
+			},
+		},
+		{
+			Name:                 "move_miners_to_rack",
+			Description:          "Atomically move explicitly identified miners to an existing rack. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"target_rack_id", "device_identifiers"},
+				"properties": map[string]any{
+					"target_rack_id": map[string]any{"type": "integer", "minimum": 1},
+					"device_identifiers": map[string]any{
+						"type": "array", "minItems": 1, "maxItems": 1000, "uniqueItems": true,
+						"items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255},
+					},
+					"force_clear_conflicting_site": map[string]any{"type": "boolean"},
+				},
+			},
+		},
+		{
+			Name:                 "set_rack_slots",
+			Description:          "Assign explicitly identified miners that already belong to a rack to specific 0-indexed row/column slot positions. Use resolve_miners with rack_ids first unless exact device identifiers are already supplied; use list_racks to convert human-facing slot numbers according to the rack layout and numbering origin. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"rack_id", "slot_assignments"},
+				"properties": map[string]any{
+					"rack_id": map[string]any{"type": "integer", "minimum": 1},
+					"slot_assignments": map[string]any{
+						"type": "array", "minItems": 1, "maxItems": maxRackSlotAssignments,
+						"items": map[string]any{
+							"type":                 "object",
+							"additionalProperties": false,
+							"required":             []string{"device_identifier", "row", "column"},
+							"properties": map[string]any{
+								"device_identifier": map[string]any{"type": "string", "minLength": 1, "maxLength": 255},
+								"row":               map[string]any{"type": "integer", "minimum": 0, "description": "0-indexed row coordinate."},
+								"column":            map[string]any{"type": "integer", "minimum": 0, "description": "0-indexed column coordinate."},
+							},
+						},
+					},
+				},
+			},
+		},
+		{
+			Name:                 "clear_rack_slots",
+			Description:          "Clear slot positions for explicitly identified miners in a rack while preserving rack membership. This write always pauses for explicit operator confirmation before execution.",
+			RequiresConfirmation: true,
+			InputSchema: map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"rack_id", "device_identifiers"},
+				"properties": map[string]any{
+					"rack_id": map[string]any{"type": "integer", "minimum": 1},
+					"device_identifiers": map[string]any{
+						"type": "array", "minItems": 1, "maxItems": maxRackSlotAssignments, "uniqueItems": true,
+						"items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255},
+					},
+				},
+			},
+		},
+	}
+}
+
+func emptyObjectSchema() map[string]any {
+	return map[string]any{"type": "object", "additionalProperties": false, "properties": map[string]any{}}
+}
+
+func minerActionSchema() map[string]any {
+	deviceIdentifiersSchema := map[string]any{
+		"type": "array", "minItems": 1, "maxItems": maxMinerActionDevices, "uniqueItems": true,
+		"items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255},
+	}
+	filterSchema := map[string]any{
+		"type":                 "object",
+		"additionalProperties": false,
+		"properties": map[string]any{
+			"device_statuses": map[string]any{
+				"type": "array",
+				"items": map[string]any{
+					"type": "string",
+					"enum": []string{
+						"online", "offline", "maintenance", "error", "inactive",
+						"needs_mining_pool", "updating", "reboot_required",
+						"hashing", "broken", "sleeping",
+					},
+				},
+				"uniqueItems": true,
+			},
+			"site_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+			"include_unassigned":  map[string]any{"type": "boolean"},
+			"building_ids":        map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+			"include_no_building": map[string]any{"type": "boolean"},
+			"rack_ids":            map[string]any{"type": "array", "items": map[string]any{"type": "integer", "minimum": 1}},
+			"include_no_rack":     map[string]any{"type": "boolean"},
+			"models":              map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 255}},
+			"ip_cidrs":            map[string]any{"type": "array", "items": map[string]any{"type": "string", "minLength": 1, "maxLength": 64}},
+		},
+	}
+	return map[string]any{
+		"type":                 "object",
+		"additionalProperties": false,
+		"required":             []string{"action"},
+		"properties": map[string]any{
+			"action": map[string]any{
+				"type": "string",
+				"enum": []string{"reboot", "start_mining", "stop_mining", "blink_led"},
+			},
+			"device_identifiers": deviceIdentifiersSchema,
+			"selector": map[string]any{
+				"type":                 "object",
+				"additionalProperties": false,
+				"required":             []string{"type"},
+				"properties": map[string]any{
+					"type":               map[string]any{"type": "string", "enum": []string{minerActionSelectorExplicit, minerActionSelectorAllDevices, minerActionSelectorFilter}},
+					"device_identifiers": deviceIdentifiersSchema,
+					"filter":             filterSchema,
+				},
+			},
+		},
+	}
+}
+
+func downtimeWindowSchema() map[string]any {
+	return map[string]any{
+		"type":                 "object",
+		"additionalProperties": false,
+		"required":             []string{"name", "action", "start_date", "start_time", "timezone", "targets"},
+		"properties": map[string]any{
+			"name":       map[string]any{"type": "string", "minLength": 1, "maxLength": 100},
+			"action":     map[string]any{"type": "string", "enum": []string{"sleep", "reboot", "set_power_target"}},
+			"start_date": map[string]any{"type": "string", "description": "YYYY-MM-DD"},
+			"start_time": map[string]any{"type": "string", "description": "HH:MM, 24-hour local time in timezone"},
+			"end_date":   map[string]any{"type": "string", "description": "Optional YYYY-MM-DD"},
+			"end_time":   map[string]any{"type": "string", "description": "Optional HH:MM. Only bounded set_power_target windows auto-revert at end_time."},
+			"timezone":   map[string]any{"type": "string", "description": "IANA timezone, for example America/Chicago"},
+			"power_target_mode": map[string]any{
+				"type":        "string",
+				"enum":        []string{"default", "max"},
+				"description": "Required only when action is set_power_target.",
+			},
+			"targets": map[string]any{
+				"type": "array", "minItems": 1, "maxItems": maxDowntimeTargets,
+				"items": map[string]any{
+					"type":                 "object",
+					"additionalProperties": false,
+					"required":             []string{"type", "target_id"},
+					"properties": map[string]any{
+						"type":      map[string]any{"type": "string", "enum": []string{"rack", "miner", "group", "site", "building"}},
+						"target_id": map[string]any{"type": "string", "minLength": 1},
+					},
+				},
+			},
+		},
+	}
+}
+
+type createSiteInput struct {
+	Name            string  `json:"name"`
+	LocationCity    string  `json:"location_city"`
+	LocationState   string  `json:"location_state"`
+	Country         string  `json:"country"`
+	Address         string  `json:"address"`
+	PostalCode      string  `json:"postal_code"`
+	Timezone        string  `json:"timezone"`
+	PowerCapacityMW float64 `json:"power_capacity_mw"`
+	Notes           string  `json:"notes"`
+}
+
+type resolveMinersInput struct {
+	Query             string   `json:"query"`
+	DeviceStatuses    []string `json:"device_statuses"`
+	SiteIDs           []int64  `json:"site_ids"`
+	IncludeUnassigned bool     `json:"include_unassigned"`
+	BuildingIDs       []int64  `json:"building_ids"`
+	IncludeNoBuilding bool     `json:"include_no_building"`
+	RackIDs           []int64  `json:"rack_ids"`
+	IncludeNoRack     bool     `json:"include_no_rack"`
+	Models            []string `json:"models"`
+	IPCIDRs           []string `json:"ip_cidrs"`
+	Limit             int32    `json:"limit"`
+}
+
+type rackHealthInput struct {
+	RackID int64 `json:"rack_id"`
+	Limit  int32 `json:"limit"`
+}
+
+type minerActionInput struct {
+	Action            string                    `json:"action"`
+	DeviceIdentifiers []string                  `json:"device_identifiers"`
+	Selector          *minerActionSelectorInput `json:"selector"`
+}
+
+type minerActionSelectorInput struct {
+	Type              string                  `json:"type"`
+	DeviceIdentifiers []string                `json:"device_identifiers"`
+	Filter            *minerActionFilterInput `json:"filter"`
+}
+
+type minerActionFilterInput struct {
+	DeviceStatuses    []string `json:"device_statuses"`
+	SiteIDs           []int64  `json:"site_ids"`
+	IncludeUnassigned bool     `json:"include_unassigned"`
+	BuildingIDs       []int64  `json:"building_ids"`
+	IncludeNoBuilding bool     `json:"include_no_building"`
+	RackIDs           []int64  `json:"rack_ids"`
+	IncludeNoRack     bool     `json:"include_no_rack"`
+	Models            []string `json:"models"`
+	IPCIDRs           []string `json:"ip_cidrs"`
+}
+
+type minerActionSelectionView struct {
+	Type              string                  `json:"type"`
+	Description       string                  `json:"description"`
+	DeviceIdentifiers []string                `json:"device_identifiers,omitempty"`
+	Filter            *minerActionFilterInput `json:"filter,omitempty"`
+}
+
+type downtimeTargetInput struct {
+	Type     string `json:"type"`
+	TargetID string `json:"target_id"`
+}
+
+type downtimeWindowInput struct {
+	Name            string                `json:"name"`
+	Action          string                `json:"action"`
+	StartDate       string                `json:"start_date"`
+	StartTime       string                `json:"start_time"`
+	EndDate         string                `json:"end_date"`
+	EndTime         string                `json:"end_time"`
+	Timezone        string                `json:"timezone"`
+	PowerTargetMode string                `json:"power_target_mode"`
+	Targets         []downtimeTargetInput `json:"targets"`
+}
+
+type createRackInput struct {
+	Label       string `json:"label"`
+	Rows        int32  `json:"rows"`
+	Columns     int32  `json:"columns"`
+	Zone        string `json:"zone"`
+	SiteID      *int64 `json:"site_id"`
+	BuildingID  *int64 `json:"building_id"`
+	OrderIndex  string `json:"order_index"`
+	CoolingType string `json:"cooling_type"`
+}
+
+type moveMinersToRackInput struct {
+	TargetRackID              int64    `json:"target_rack_id"`
+	DeviceIdentifiers         []string `json:"device_identifiers"`
+	ForceClearConflictingSite bool     `json:"force_clear_conflicting_site"`
+}
+
+type rackSlotAssignmentInput struct {
+	DeviceIdentifier string `json:"device_identifier"`
+	Row              int32  `json:"row"`
+	Column           int32  `json:"column"`
+}
+
+type setRackSlotsInput struct {
+	RackID          int64                     `json:"rack_id"`
+	SlotAssignments []rackSlotAssignmentInput `json:"slot_assignments"`
+}
+
+type clearRackSlotsInput struct {
+	RackID            int64    `json:"rack_id"`
+	DeviceIdentifiers []string `json:"device_identifiers"`
+}
+
+func decodeToolArguments(arguments json.RawMessage, destination any) error {
+	if len(arguments) == 0 {
+		arguments = json.RawMessage(`{}`)
+	}
+	decoder := json.NewDecoder(bytes.NewReader(arguments))
+	decoder.DisallowUnknownFields()
+	if err := decoder.Decode(destination); err != nil {
+		return fmt.Errorf("decode tool arguments: %w", err)
+	}
+	if err := decoder.Decode(&struct{}{}); err != io.EOF {
+		if err == nil {
+			return fmt.Errorf("multiple JSON values are not allowed")
+		}
+		return fmt.Errorf("check for trailing JSON values: %w", err)
+	}
+	return nil
+}
+
+func buildCreateSiteRequest(arguments json.RawMessage) (createSiteInput, *sitesv1.CreateSiteRequest, error) {
+	var input createSiteInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, nil, fmt.Errorf("invalid create_site arguments: %w", err)
+	}
+	request := &sitesv1.CreateSiteRequest{
+		Name:            input.Name,
+		LocationCity:    input.LocationCity,
+		LocationState:   input.LocationState,
+		Country:         input.Country,
+		Address:         input.Address,
+		PostalCode:      input.PostalCode,
+		Timezone:        input.Timezone,
+		PowerCapacityMw: input.PowerCapacityMW,
+		Notes:           input.Notes,
+	}
+	if err := protovalidate.Validate(request); err != nil {
+		return input, nil, fmt.Errorf("invalid create_site arguments: %w", err)
+	}
+	return input, request, nil
+}
+
+func buildResolveMinersFilter(arguments json.RawMessage) (resolveMinersInput, *fleetv1.MinerListFilter, error) {
+	var input resolveMinersInput
+	if len(arguments) != 0 {
+		if err := decodeToolArguments(arguments, &input); err != nil {
+			return input, nil, fmt.Errorf("invalid resolve_miners arguments: %w", err)
+		}
+	}
+	input.Query = strings.TrimSpace(input.Query)
+	if len(input.Query) > 255 {
+		return input, nil, fmt.Errorf("invalid resolve_miners arguments: query must be 255 characters or fewer")
+	}
+	if input.Limit == 0 {
+		input.Limit = defaultResolveMinersLimit
+	}
+	if input.Limit < 1 || input.Limit > maxResolveMinersLimit {
+		return input, nil, fmt.Errorf("invalid resolve_miners arguments: limit must be between 1 and %d", maxResolveMinersLimit)
+	}
+	if err := validatePositiveIDs("site_ids", input.SiteIDs); err != nil {
+		return input, nil, err
+	}
+	if err := validatePositiveIDs("building_ids", input.BuildingIDs); err != nil {
+		return input, nil, err
+	}
+	if err := validatePositiveIDs("rack_ids", input.RackIDs); err != nil {
+		return input, nil, err
+	}
+	statuses := make([]fleetv1.DeviceStatus, 0, len(input.DeviceStatuses))
+	for _, status := range input.DeviceStatuses {
+		parsed, err := parseResolveMinerStatus(status)
+		if err != nil {
+			return input, nil, err
+		}
+		statuses = append(statuses, parsed)
+	}
+	filter := &fleetv1.MinerListFilter{
+		DeviceStatus:      statuses,
+		SiteIds:           input.SiteIDs,
+		IncludeUnassigned: input.IncludeUnassigned,
+		BuildingIds:       input.BuildingIDs,
+		IncludeNoBuilding: input.IncludeNoBuilding,
+		RackIds:           input.RackIDs,
+		IncludeNoRack:     input.IncludeNoRack,
+		Models:            input.Models,
+		IpCidrs:           input.IPCIDRs,
+	}
+	return input, filter, nil
+}
+
+func validatePositiveIDs(field string, ids []int64) error {
+	return validatePositiveIDsForTool("resolve_miners", field, ids)
+}
+
+func validatePositiveIDsForTool(toolName, field string, ids []int64) error {
+	for _, id := range ids {
+		if id <= 0 {
+			return fmt.Errorf("invalid %s arguments: %s must contain only positive IDs", toolName, field)
+		}
+	}
+	return nil
+}
+
+func parseResolveMinerStatus(status string) (fleetv1.DeviceStatus, error) {
+	return parseDeviceStatusArgument("resolve_miners", status)
+}
+
+func parseDeviceStatusArgument(toolName, status string) (fleetv1.DeviceStatus, error) {
+	normalized := strings.ToLower(strings.TrimSpace(status))
+	normalized = strings.ReplaceAll(normalized, "-", "_")
+	normalized = strings.ReplaceAll(normalized, " ", "_")
+	switch normalized {
+	case "online", "hashing":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_ONLINE, nil
+	case "offline":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE, nil
+	case "maintenance":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_MAINTENANCE, nil
+	case "error", "broken":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_ERROR, nil
+	case "inactive", "sleeping":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_INACTIVE, nil
+	case "needs_mining_pool", "needs_pool", "pool_needed":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_NEEDS_MINING_POOL, nil
+	case "updating":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_UPDATING, nil
+	case "reboot_required":
+		return fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED, nil
+	default:
+		return fleetv1.DeviceStatus_DEVICE_STATUS_UNSPECIFIED, fmt.Errorf("invalid %s arguments: unsupported device_status %q", toolName, status)
+	}
+}
+
+func buildCreateRackRequest(arguments json.RawMessage) (createRackInput, *devicesetv1.CreateDeviceSetRequest, error) {
+	var input createRackInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, nil, fmt.Errorf("invalid create_rack arguments: %w", err)
+	}
+	if input.SiteID != nil && input.BuildingID != nil {
+		return input, nil, fmt.Errorf("invalid create_rack arguments: specify site_id or building_id, not both")
+	}
+	if input.SiteID != nil && *input.SiteID <= 0 {
+		return input, nil, fmt.Errorf("invalid create_rack arguments: site_id must be positive")
+	}
+	if input.BuildingID != nil && *input.BuildingID <= 0 {
+		return input, nil, fmt.Errorf("invalid create_rack arguments: building_id must be positive")
+	}
+	orderIndex := devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_TOP_LEFT
+	switch input.OrderIndex {
+	case "", rackOrderIndexTopLeft:
+	case rackOrderIndexTopRight:
+		orderIndex = devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_TOP_RIGHT
+	case rackOrderIndexBottomLeft:
+		orderIndex = devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_BOTTOM_LEFT
+	case rackOrderIndexBottomRight:
+		orderIndex = devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_BOTTOM_RIGHT
+	default:
+		return input, nil, fmt.Errorf("invalid create_rack order_index %q", input.OrderIndex)
+	}
+	coolingType := devicesetv1.RackCoolingType_RACK_COOLING_TYPE_AIR
+	switch input.CoolingType {
+	case "", "air":
+	case "immersion":
+		coolingType = devicesetv1.RackCoolingType_RACK_COOLING_TYPE_IMMERSION
+	default:
+		return input, nil, fmt.Errorf("invalid create_rack cooling_type %q", input.CoolingType)
+	}
+	request := &devicesetv1.CreateDeviceSetRequest{
+		Type:  devicesetv1.DeviceSetType_DEVICE_SET_TYPE_RACK,
+		Label: input.Label,
+		TypeDetails: &devicesetv1.CreateDeviceSetRequest_RackInfo{RackInfo: &devicesetv1.RackInfo{
+			Rows:        input.Rows,
+			Columns:     input.Columns,
+			Zone:        input.Zone,
+			OrderIndex:  orderIndex,
+			CoolingType: coolingType,
+			SiteId:      input.SiteID,
+			BuildingId:  input.BuildingID,
+		}},
+	}
+	if err := protovalidate.Validate(request); err != nil {
+		return input, nil, fmt.Errorf("invalid create_rack arguments: %w", err)
+	}
+	return input, request, nil
+}
+
+func buildMoveMinersRequest(arguments json.RawMessage) (moveMinersToRackInput, *devicesetv1.AssignDevicesToRackRequest, error) {
+	var input moveMinersToRackInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, nil, fmt.Errorf("invalid move_miners_to_rack arguments: %w", err)
+	}
+	if len(input.DeviceIdentifiers) == 0 || len(input.DeviceIdentifiers) > 1000 {
+		return input, nil, fmt.Errorf("invalid move_miners_to_rack arguments: device_identifiers must contain between 1 and 1000 miners")
+	}
+	seen := make(map[string]struct{}, len(input.DeviceIdentifiers))
+	for _, identifier := range input.DeviceIdentifiers {
+		if strings.TrimSpace(identifier) == "" || len(identifier) > 255 {
+			return input, nil, fmt.Errorf("invalid move_miners_to_rack arguments: device identifiers must contain 1 to 255 characters")
+		}
+		if _, exists := seen[identifier]; exists {
+			return input, nil, fmt.Errorf("invalid move_miners_to_rack arguments: duplicate device identifier %q", identifier)
+		}
+		seen[identifier] = struct{}{}
+	}
+	force := input.ForceClearConflictingSite
+	request := &devicesetv1.AssignDevicesToRackRequest{
+		TargetRackId: &input.TargetRackID,
+		DeviceSelector: &commonv1.DeviceSelector{
+			SelectionType: &commonv1.DeviceSelector_DeviceList{DeviceList: &commonv1.DeviceIdentifierList{
+				DeviceIdentifiers: input.DeviceIdentifiers,
+			}},
+		},
+		ForceClearConflictingSite: &force,
+	}
+	if err := protovalidate.Validate(request); err != nil {
+		return input, nil, fmt.Errorf("invalid move_miners_to_rack arguments: %w", err)
+	}
+	return input, request, nil
+}
+
+func buildGetRackSlotsRequest(arguments json.RawMessage) (int64, *devicesetv1.GetRackSlotsRequest, error) {
+	var input struct {
+		RackID int64 `json:"rack_id"`
+	}
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return 0, nil, fmt.Errorf("invalid get_rack_slots arguments: %w", err)
+	}
+	if input.RackID <= 0 {
+		return 0, nil, fmt.Errorf("invalid get_rack_slots arguments: rack_id must be positive")
+	}
+	request := &devicesetv1.GetRackSlotsRequest{DeviceSetId: input.RackID}
+	if err := protovalidate.Validate(request); err != nil {
+		return 0, nil, fmt.Errorf("invalid get_rack_slots arguments: %w", err)
+	}
+	return input.RackID, request, nil
+}
+
+func buildRackHealthInput(arguments json.RawMessage) (rackHealthInput, error) {
+	var input rackHealthInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, fmt.Errorf("invalid get_rack_health arguments: %w", err)
+	}
+	if input.RackID <= 0 {
+		return input, fmt.Errorf("invalid get_rack_health arguments: rack_id must be positive")
+	}
+	if input.Limit == 0 {
+		input.Limit = maxResolveMinersLimit
+	}
+	if input.Limit < 1 || input.Limit > maxResolveMinersLimit {
+		return input, fmt.Errorf("invalid get_rack_health arguments: limit must be between 1 and %d", maxResolveMinersLimit)
+	}
+	return input, nil
+}
+
+func buildMinerActionInput(arguments json.RawMessage, toolName string) (minerActionInput, minercommandv1.CommandType, *minercommandv1.DeviceSelector, minerActionSelectionView, error) {
+	var input minerActionInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, minercommandv1.CommandType_COMMAND_TYPE_UNSPECIFIED, nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: %w", toolName, err)
+	}
+	commandType, err := parseMinerActionCommandType(input.Action)
+	if err != nil {
+		return input, minercommandv1.CommandType_COMMAND_TYPE_UNSPECIFIED, nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: %w", toolName, err)
+	}
+	selector, selection, err := buildMinerActionSelector(input, toolName)
+	if err != nil {
+		return input, minercommandv1.CommandType_COMMAND_TYPE_UNSPECIFIED, nil, minerActionSelectionView{}, err
+	}
+	return input, commandType, selector, selection, nil
+}
+
+func parseMinerActionCommandType(action string) (minercommandv1.CommandType, error) {
+	switch normalizeToolEnum(action) {
+	case "reboot":
+		return minercommandv1.CommandType_COMMAND_TYPE_REBOOT, nil
+	case "start_mining":
+		return minercommandv1.CommandType_COMMAND_TYPE_START_MINING, nil
+	case "stop_mining":
+		return minercommandv1.CommandType_COMMAND_TYPE_STOP_MINING, nil
+	case "blink_led":
+		return minercommandv1.CommandType_COMMAND_TYPE_BLINK_LED, nil
+	default:
+		return minercommandv1.CommandType_COMMAND_TYPE_UNSPECIFIED, fmt.Errorf("unsupported action %q", action)
+	}
+}
+
+func commandSelectorForDevices(deviceIdentifiers []string) *minercommandv1.DeviceSelector {
+	return &minercommandv1.DeviceSelector{
+		SelectionType: &minercommandv1.DeviceSelector_IncludeDevices{
+			IncludeDevices: &commonv1.DeviceIdentifierList{
+				DeviceIdentifiers: deviceIdentifiers,
+			},
+		},
+	}
+}
+
+func buildMinerActionSelector(input minerActionInput, toolName string) (*minercommandv1.DeviceSelector, minerActionSelectionView, error) {
+	if input.Selector == nil {
+		if err := validateExplicitDeviceIdentifiers(toolName, input.DeviceIdentifiers, maxMinerActionDevices); err != nil {
+			return nil, minerActionSelectionView{}, err
+		}
+		return commandSelectorForDevices(input.DeviceIdentifiers), minerActionSelectionView{
+			Type:              minerActionSelectorExplicit,
+			Description:       fmt.Sprintf("%d explicit miner(s)", len(input.DeviceIdentifiers)),
+			DeviceIdentifiers: input.DeviceIdentifiers,
+		}, nil
+	}
+	if len(input.DeviceIdentifiers) > 0 {
+		return nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: specify either top-level device_identifiers or selector, not both", toolName)
+	}
+	switch normalizeToolEnum(input.Selector.Type) {
+	case minerActionSelectorExplicit:
+		if err := validateExplicitDeviceIdentifiers(toolName, input.Selector.DeviceIdentifiers, maxMinerActionDevices); err != nil {
+			return nil, minerActionSelectionView{}, err
+		}
+		if input.Selector.Filter != nil {
+			return nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: selector.filter is only valid when selector.type is filter", toolName)
+		}
+		return commandSelectorForDevices(input.Selector.DeviceIdentifiers), minerActionSelectionView{
+			Type:              minerActionSelectorExplicit,
+			Description:       fmt.Sprintf("%d explicit miner(s)", len(input.Selector.DeviceIdentifiers)),
+			DeviceIdentifiers: input.Selector.DeviceIdentifiers,
+		}, nil
+	case minerActionSelectorAllDevices:
+		if len(input.Selector.DeviceIdentifiers) > 0 || input.Selector.Filter != nil {
+			return nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: all_devices selector does not accept device_identifiers or filter", toolName)
+		}
+		return &minercommandv1.DeviceSelector{
+				SelectionType: &minercommandv1.DeviceSelector_AllDevices{
+					AllDevices: &minercommandv1.DeviceFilter{},
+				},
+			}, minerActionSelectionView{
+				Type:        minerActionSelectorAllDevices,
+				Description: "whole fleet command-eligible miners",
+			}, nil
+	case minerActionSelectorFilter:
+		if len(input.Selector.DeviceIdentifiers) > 0 {
+			return nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: selector.device_identifiers is only valid when selector.type is explicit", toolName)
+		}
+		filter, err := buildMinerActionFilter(input.Selector.Filter, toolName)
+		if err != nil {
+			return nil, minerActionSelectionView{}, err
+		}
+		filterInput := input.Selector.Filter
+		if filterInput == nil {
+			filterInput = &minerActionFilterInput{}
+		}
+		return &minercommandv1.DeviceSelector{
+				SelectionType: &minercommandv1.DeviceSelector_AllMatchingFilter{
+					AllMatchingFilter: filter,
+				},
+			}, minerActionSelectionView{
+				Type:        minerActionSelectorFilter,
+				Description: "miners matching the supplied filter",
+				Filter:      filterInput,
+			}, nil
+	default:
+		return nil, minerActionSelectionView{}, fmt.Errorf("invalid %s arguments: unsupported selector.type %q", toolName, input.Selector.Type)
+	}
+}
+
+func buildMinerActionFilter(input *minerActionFilterInput, toolName string) (*fleetv1.MinerListFilter, error) {
+	if input == nil {
+		return &fleetv1.MinerListFilter{}, nil
+	}
+	if err := validatePositiveIDsForTool(toolName, "site_ids", input.SiteIDs); err != nil {
+		return nil, err
+	}
+	if err := validatePositiveIDsForTool(toolName, "building_ids", input.BuildingIDs); err != nil {
+		return nil, err
+	}
+	if err := validatePositiveIDsForTool(toolName, "rack_ids", input.RackIDs); err != nil {
+		return nil, err
+	}
+	statuses := make([]fleetv1.DeviceStatus, 0, len(input.DeviceStatuses))
+	for _, status := range input.DeviceStatuses {
+		parsed, err := parseDeviceStatusArgument(toolName, status)
+		if err != nil {
+			return nil, err
+		}
+		statuses = append(statuses, parsed)
+	}
+	return &fleetv1.MinerListFilter{
+		DeviceStatus:      statuses,
+		SiteIds:           input.SiteIDs,
+		IncludeUnassigned: input.IncludeUnassigned,
+		BuildingIds:       input.BuildingIDs,
+		IncludeNoBuilding: input.IncludeNoBuilding,
+		RackIds:           input.RackIDs,
+		IncludeNoRack:     input.IncludeNoRack,
+		Models:            input.Models,
+		IpCidrs:           input.IPCIDRs,
+	}, nil
+}
+
+func validateExplicitDeviceIdentifiers(toolName string, identifiers []string, maxItems int) error {
+	if len(identifiers) == 0 || len(identifiers) > maxItems {
+		return fmt.Errorf("invalid %s arguments: device_identifiers must contain between 1 and %d miners", toolName, maxItems)
+	}
+	seen := make(map[string]struct{}, len(identifiers))
+	for _, identifier := range identifiers {
+		if strings.TrimSpace(identifier) == "" || len(identifier) > 255 {
+			return fmt.Errorf("invalid %s arguments: device identifiers must contain 1 to 255 characters", toolName)
+		}
+		if _, exists := seen[identifier]; exists {
+			return fmt.Errorf("invalid %s arguments: duplicate device identifier %q", toolName, identifier)
+		}
+		seen[identifier] = struct{}{}
+	}
+	return nil
+}
+
+func buildDowntimeWindowRequest(arguments json.RawMessage) (downtimeWindowInput, *schedulev1.CreateScheduleRequest, error) {
+	var input downtimeWindowInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, nil, fmt.Errorf("invalid downtime window arguments: %w", err)
+	}
+	action, err := parseScheduleAction(input.Action)
+	if err != nil {
+		return input, nil, fmt.Errorf("invalid downtime window arguments: %w", err)
+	}
+	if len(input.Targets) == 0 || len(input.Targets) > maxDowntimeTargets {
+		return input, nil, fmt.Errorf("invalid downtime window arguments: targets must contain between 1 and %d items", maxDowntimeTargets)
+	}
+	targets := make([]*schedulev1.ScheduleTarget, 0, len(input.Targets))
+	for i, target := range input.Targets {
+		targetType, err := parseScheduleTargetType(target.Type)
+		if err != nil {
+			return input, nil, fmt.Errorf("invalid downtime window arguments: targets[%d]: %w", i, err)
+		}
+		targetID := strings.TrimSpace(target.TargetID)
+		if targetID == "" {
+			return input, nil, fmt.Errorf("invalid downtime window arguments: targets[%d].target_id is required", i)
+		}
+		targets = append(targets, &schedulev1.ScheduleTarget{TargetType: targetType, TargetId: targetID})
+	}
+	var actionConfig *schedulev1.PowerTargetConfig
+	if action == schedulev1.ScheduleAction_SCHEDULE_ACTION_SET_POWER_TARGET {
+		mode, err := parsePowerTargetMode(input.PowerTargetMode)
+		if err != nil {
+			return input, nil, fmt.Errorf("invalid downtime window arguments: %w", err)
+		}
+		actionConfig = &schedulev1.PowerTargetConfig{Mode: mode}
+	} else if strings.TrimSpace(input.PowerTargetMode) != "" {
+		return input, nil, fmt.Errorf("invalid downtime window arguments: power_target_mode is only valid for set_power_target")
+	}
+	request := &schedulev1.CreateScheduleRequest{
+		Name:         input.Name,
+		Action:       action,
+		ActionConfig: actionConfig,
+		ScheduleType: schedulev1.ScheduleType_SCHEDULE_TYPE_ONE_TIME,
+		StartDate:    input.StartDate,
+		StartTime:    input.StartTime,
+		EndDate:      input.EndDate,
+		EndTime:      input.EndTime,
+		Timezone:     input.Timezone,
+		Targets:      targets,
+	}
+	if err := protovalidate.Validate(request); err != nil {
+		return input, nil, fmt.Errorf("invalid downtime window arguments: %w", err)
+	}
+	return input, request, nil
+}
+
+func parseScheduleAction(action string) (schedulev1.ScheduleAction, error) {
+	switch normalizeToolEnum(action) {
+	case "sleep":
+		return schedulev1.ScheduleAction_SCHEDULE_ACTION_SLEEP, nil
+	case "reboot":
+		return schedulev1.ScheduleAction_SCHEDULE_ACTION_REBOOT, nil
+	case "set_power_target":
+		return schedulev1.ScheduleAction_SCHEDULE_ACTION_SET_POWER_TARGET, nil
+	default:
+		return schedulev1.ScheduleAction_SCHEDULE_ACTION_UNSPECIFIED, fmt.Errorf("unsupported action %q", action)
+	}
+}
+
+func parsePowerTargetMode(mode string) (schedulev1.PowerTargetMode, error) {
+	switch normalizeToolEnum(mode) {
+	case "default":
+		return schedulev1.PowerTargetMode_POWER_TARGET_MODE_DEFAULT, nil
+	case "max":
+		return schedulev1.PowerTargetMode_POWER_TARGET_MODE_MAX, nil
+	default:
+		return schedulev1.PowerTargetMode_POWER_TARGET_MODE_UNSPECIFIED, fmt.Errorf("power_target_mode must be default or max")
+	}
+}
+
+func parseScheduleTargetType(targetType string) (schedulev1.ScheduleTargetType, error) {
+	switch normalizeToolEnum(targetType) {
+	case "rack":
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_RACK, nil
+	case "miner":
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_MINER, nil
+	case "group":
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_GROUP, nil
+	case "site":
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_SITE, nil
+	case "building":
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_BUILDING, nil
+	default:
+		return schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_UNSPECIFIED, fmt.Errorf("unsupported target type %q", targetType)
+	}
+}
+
+func normalizeToolEnum(value string) string {
+	normalized := strings.ToLower(strings.TrimSpace(value))
+	normalized = strings.ReplaceAll(normalized, "-", "_")
+	normalized = strings.ReplaceAll(normalized, " ", "_")
+	return normalized
+}
+
+func buildSetRackSlotsInput(arguments json.RawMessage) (setRackSlotsInput, error) {
+	var input setRackSlotsInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, fmt.Errorf("invalid set_rack_slots arguments: %w", err)
+	}
+	if input.RackID <= 0 {
+		return input, fmt.Errorf("invalid set_rack_slots arguments: rack_id must be positive")
+	}
+	if len(input.SlotAssignments) == 0 || len(input.SlotAssignments) > maxRackSlotAssignments {
+		return input, fmt.Errorf("invalid set_rack_slots arguments: slot_assignments must contain between 1 and %d assignments", maxRackSlotAssignments)
+	}
+	seenDevices := make(map[string]struct{}, len(input.SlotAssignments))
+	seenPositions := make(map[string]string, len(input.SlotAssignments))
+	for i, assignment := range input.SlotAssignments {
+		if strings.TrimSpace(assignment.DeviceIdentifier) == "" || len(assignment.DeviceIdentifier) > 255 {
+			return input, fmt.Errorf("invalid set_rack_slots arguments: slot_assignments[%d].device_identifier must contain 1 to 255 characters", i)
+		}
+		if _, exists := seenDevices[assignment.DeviceIdentifier]; exists {
+			return input, fmt.Errorf("invalid set_rack_slots arguments: duplicate device identifier %q", assignment.DeviceIdentifier)
+		}
+		seenDevices[assignment.DeviceIdentifier] = struct{}{}
+		if assignment.Row < 0 || assignment.Column < 0 {
+			return input, fmt.Errorf("invalid set_rack_slots arguments: row and column must be 0-indexed non-negative coordinates")
+		}
+		positionKey := rackSlotPositionKey(assignment.Row, assignment.Column)
+		if existing, exists := seenPositions[positionKey]; exists {
+			return input, fmt.Errorf("invalid set_rack_slots arguments: devices %q and %q target the same slot (%d,%d)", existing, assignment.DeviceIdentifier, assignment.Row, assignment.Column)
+		}
+		seenPositions[positionKey] = assignment.DeviceIdentifier
+	}
+	return input, nil
+}
+
+func buildClearRackSlotsInput(arguments json.RawMessage) (clearRackSlotsInput, error) {
+	var input clearRackSlotsInput
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return input, fmt.Errorf("invalid clear_rack_slots arguments: %w", err)
+	}
+	if input.RackID <= 0 {
+		return input, fmt.Errorf("invalid clear_rack_slots arguments: rack_id must be positive")
+	}
+	if len(input.DeviceIdentifiers) == 0 || len(input.DeviceIdentifiers) > maxRackSlotAssignments {
+		return input, fmt.Errorf("invalid clear_rack_slots arguments: device_identifiers must contain between 1 and %d miners", maxRackSlotAssignments)
+	}
+	seen := make(map[string]struct{}, len(input.DeviceIdentifiers))
+	for _, identifier := range input.DeviceIdentifiers {
+		if strings.TrimSpace(identifier) == "" || len(identifier) > 255 {
+			return input, fmt.Errorf("invalid clear_rack_slots arguments: device identifiers must contain 1 to 255 characters")
+		}
+		if _, exists := seen[identifier]; exists {
+			return input, fmt.Errorf("invalid clear_rack_slots arguments: duplicate device identifier %q", identifier)
+		}
+		seen[identifier] = struct{}{}
+	}
+	return input, nil
+}
+
+func (t *FleetTools) Confirmation(name string, arguments json.RawMessage) (*chatdomain.ToolConfirmation, error) {
+	switch name {
+	case "create_site":
+		input, _, err := buildCreateSiteRequest(arguments)
+		if err != nil {
+			return nil, err
+		}
+		details := []chatdomain.ToolConfirmationDetail{{Label: "Name", Value: input.Name}}
+		locationParts := make([]string, 0, 3)
+		for _, part := range []string{input.Address, input.LocationCity, input.LocationState, input.PostalCode, input.Country} {
+			if strings.TrimSpace(part) != "" {
+				locationParts = append(locationParts, part)
+			}
+		}
+		location := strings.Join(locationParts, ", ")
+		if location != "" {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Location", Value: location})
+		}
+		if input.PowerCapacityMW > 0 {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Power capacity", Value: fmt.Sprintf("%g MW", input.PowerCapacityMW)})
+		}
+		if input.Timezone != "" {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Timezone", Value: input.Timezone})
+		}
+		if input.Notes != "" {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Notes", Value: input.Notes})
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        "Create this site?",
+			Description:  "Proto AI will add a new site to your fleet.",
+			ConfirmLabel: "Create site",
+			Details:      details,
+		}, nil
+	case "create_rack":
+		input, _, err := buildCreateRackRequest(arguments)
+		if err != nil {
+			return nil, err
+		}
+		details := []chatdomain.ToolConfirmationDetail{
+			{Label: "Rack", Value: input.Label},
+			{Label: "Layout", Value: fmt.Sprintf("%d rows × %d columns", input.Rows, input.Columns)},
+		}
+		if input.Zone != "" {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Zone", Value: input.Zone})
+		}
+		orderIndex := input.OrderIndex
+		if orderIndex == "" {
+			orderIndex = rackOrderIndexTopLeft
+		}
+		coolingType := input.CoolingType
+		if coolingType == "" {
+			coolingType = "air"
+		}
+		details = append(details,
+			chatdomain.ToolConfirmationDetail{Label: "Numbering starts", Value: strings.ReplaceAll(orderIndex, "_", " ")},
+			chatdomain.ToolConfirmationDetail{Label: "Cooling", Value: coolingType},
+		)
+		if input.BuildingID != nil {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Building ID", Value: fmt.Sprint(*input.BuildingID)})
+		} else if input.SiteID != nil {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Site ID", Value: fmt.Sprint(*input.SiteID)})
+		} else {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Placement", Value: "Unassigned"})
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        "Create this rack?",
+			Description:  "Proto AI will create an empty rack with this layout and placement.",
+			ConfirmLabel: "Create rack",
+			Details:      details,
+		}, nil
+	case "move_miners_to_rack":
+		input, _, err := buildMoveMinersRequest(arguments)
+		if err != nil {
+			return nil, err
+		}
+		description := "Proto AI will atomically replace the rack membership for these miners."
+		if input.ForceClearConflictingSite {
+			description += " Miners assigned to a site may have their site and building cleared if the destination rack is unassigned."
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        fmt.Sprintf("Move %d miner(s)?", len(input.DeviceIdentifiers)),
+			Description:  description,
+			ConfirmLabel: "Move miners",
+			Details: []chatdomain.ToolConfirmationDetail{
+				{Label: "Destination rack ID", Value: fmt.Sprint(input.TargetRackID)},
+				{Label: "Miners", Value: strings.Join(input.DeviceIdentifiers, ", ")},
+			},
+		}, nil
+	case "set_rack_slots":
+		input, err := buildSetRackSlotsInput(arguments)
+		if err != nil {
+			return nil, err
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        fmt.Sprintf("Assign %d rack slot(s)?", len(input.SlotAssignments)),
+			Description:  "Proto AI will update slot positions for miners already assigned to this rack. Rack membership and rack placement are preserved.",
+			ConfirmLabel: "Assign slots",
+			Details: []chatdomain.ToolConfirmationDetail{
+				{Label: "Rack ID", Value: fmt.Sprint(input.RackID)},
+				{Label: "Slots", Value: formatRackSlotAssignments(input.SlotAssignments, 20)},
+			},
+		}, nil
+	case "clear_rack_slots":
+		input, err := buildClearRackSlotsInput(arguments)
+		if err != nil {
+			return nil, err
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        fmt.Sprintf("Clear %d rack slot(s)?", len(input.DeviceIdentifiers)),
+			Description:  "Proto AI will clear slot positions for these miners while preserving rack membership.",
+			ConfirmLabel: "Clear slots",
+			Details: []chatdomain.ToolConfirmationDetail{
+				{Label: "Rack ID", Value: fmt.Sprint(input.RackID)},
+				{Label: "Miners", Value: formatStringList(input.DeviceIdentifiers, 20)},
+			},
+		}, nil
+	case "execute_miner_action":
+		input, _, _, selection, err := buildMinerActionInput(arguments, "execute_miner_action")
+		if err != nil {
+			return nil, err
+		}
+		actionLabel := strings.ReplaceAll(normalizeToolEnum(input.Action), "_", " ")
+		return &chatdomain.ToolConfirmation{
+			Title:        fmt.Sprintf("Execute %s on %s?", actionLabel, minerActionTargetPhrase(selection)),
+			Description:  "Proto AI will submit this command through the existing miner command service.",
+			ConfirmLabel: "Execute command",
+			Details: []chatdomain.ToolConfirmationDetail{
+				{Label: "Action", Value: actionLabel},
+				{Label: "Selection", Value: minerActionSelectionDetail(selection)},
+			},
+		}, nil
+	case "create_downtime_window":
+		input, request, err := buildDowntimeWindowRequest(arguments)
+		if err != nil {
+			return nil, err
+		}
+		details := []chatdomain.ToolConfirmationDetail{
+			{Label: "Name", Value: input.Name},
+			{Label: "Action", Value: scheduleActionLabel(request.GetAction())},
+			{Label: "Starts", Value: fmt.Sprintf("%s %s %s", input.StartDate, input.StartTime, input.Timezone)},
+			{Label: "Targets", Value: formatDowntimeTargets(input.Targets, 20)},
+		}
+		if input.EndTime != "" {
+			endDate := input.EndDate
+			if endDate == "" {
+				endDate = input.StartDate
+			}
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Ends", Value: fmt.Sprintf("%s %s %s", endDate, input.EndTime, input.Timezone)})
+		}
+		if request.GetAction() == schedulev1.ScheduleAction_SCHEDULE_ACTION_SLEEP {
+			details = append(details, chatdomain.ToolConfirmationDetail{Label: "Resume", Value: "Not automatic for sleep schedules"})
+		}
+		return &chatdomain.ToolConfirmation{
+			Title:        fmt.Sprintf("Create schedule %q?", input.Name),
+			Description:  "Proto AI will create this one-time maintenance schedule through the existing schedule service.",
+			ConfirmLabel: "Create schedule",
+			Details:      details,
+		}, nil
+	default:
+		return nil, nil
+	}
+}
+
+func (t *FleetTools) Execute(ctx context.Context, name string, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	switch name {
+	case "get_miner_state_counts":
+		return t.getMinerStateCounts(ctx, arguments)
+	case "resolve_miners":
+		return t.resolveMiners(ctx, arguments)
+	case "list_sites":
+		return t.listSites(ctx)
+	case "get_site_health_summary":
+		return t.getSiteHealthSummary(ctx, arguments)
+	case "list_pools":
+		return t.listPools(ctx)
+	case "list_actionable_miner_issues":
+		return t.listActionableMinerIssues(ctx, arguments)
+	case "list_racks":
+		return t.listRacks(ctx)
+	case "get_rack_slots":
+		return t.getRackSlots(ctx, arguments)
+	case "get_rack_health":
+		return t.getRackHealth(ctx, arguments)
+	case "preview_miner_action":
+		return t.previewMinerAction(ctx, arguments)
+	case "execute_miner_action":
+		return t.executeMinerAction(ctx, arguments)
+	case "preview_downtime_window":
+		return t.previewDowntimeWindow(ctx, arguments)
+	case "create_downtime_window":
+		return t.createDowntimeWindow(ctx, arguments)
+	case "create_site":
+		return t.createSite(ctx, arguments)
+	case "create_rack":
+		return t.createRack(ctx, arguments)
+	case "move_miners_to_rack":
+		return t.moveMinersToRack(ctx, arguments)
+	case "set_rack_slots":
+		return t.setRackSlots(ctx, arguments)
+	case "clear_rack_slots":
+		return t.clearRackSlots(ctx, arguments)
+	default:
+		return chatdomain.ToolOutput{}, fmt.Errorf("unknown tool %q", name)
+	}
+}
+
+func (t *FleetTools) getMinerStateCounts(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	var input struct {
+		SiteIDs           []int64 `json:"site_ids"`
+		IncludeUnassigned bool    `json:"include_unassigned"`
+	}
+	if len(arguments) != 0 {
+		decoder := json.NewDecoder(bytes.NewReader(arguments))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&input); err != nil {
+			return chatdomain.ToolOutput{}, fmt.Errorf("invalid get_miner_state_counts arguments: %w", err)
+		}
+	}
+	response, err := t.fleet.GetMinerStateCounts(ctx, connect.NewRequest(&fleetv1.GetMinerStateCountsRequest{
+		SiteIds:           input.SiteIDs,
+		IncludeUnassigned: input.IncludeUnassigned,
+	}))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	counts := response.Msg.GetStateCounts()
+	payload := map[string]any{
+		"total_miners": response.Msg.GetTotalMiners(),
+		"hashing":      counts.GetHashingCount(),
+		"broken":       counts.GetBrokenCount(),
+		"offline":      counts.GetOfflineCount(),
+		"sleeping":     counts.GetSleepingCount(),
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal miner state counts: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Read state for %d miners", response.Msg.GetTotalMiners()),
+	}, nil
+}
+
+func (t *FleetTools) getSiteHealthSummary(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	var input struct {
+		SiteID            *int64 `json:"site_id"`
+		IncludeUnassigned bool   `json:"include_unassigned"`
+	}
+	if err := decodeToolArguments(arguments, &input); err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("invalid get_site_health_summary arguments: %w", err)
+	}
+	siteIDs := []int64(nil)
+	if input.SiteID != nil {
+		if *input.SiteID <= 0 {
+			return chatdomain.ToolOutput{}, fmt.Errorf("invalid get_site_health_summary arguments: site_id must be positive")
+		}
+		siteIDs = []int64{*input.SiteID}
+	}
+	response, err := t.fleet.GetMinerStateCounts(ctx, connect.NewRequest(&fleetv1.GetMinerStateCountsRequest{
+		SiteIds:           siteIDs,
+		IncludeUnassigned: input.IncludeUnassigned,
+	}))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	counts := response.Msg.GetStateCounts()
+
+	siteInventory, err := t.siteInventory(ctx, input.SiteID)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	payload := map[string]any{
+		"scope": map[string]any{
+			"site_id":            input.SiteID,
+			"include_unassigned": input.IncludeUnassigned,
+		},
+		"total_miners": response.Msg.GetTotalMiners(),
+		"state_counts": map[string]any{
+			"hashing":  counts.GetHashingCount(),
+			"broken":   counts.GetBrokenCount(),
+			"offline":  counts.GetOfflineCount(),
+			"sleeping": counts.GetSleepingCount(),
+		},
+		"sites": siteInventory,
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal site health summary: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Read health summary for %d miner(s)", response.Msg.GetTotalMiners()),
+	}, nil
+}
+
+type resourceRefView struct {
+	ID    int64  `json:"id"`
+	Label string `json:"label"`
+}
+
+type resolvedMinerView struct {
+	DeviceIdentifier string           `json:"device_identifier"`
+	Name             string           `json:"name,omitempty"`
+	Status           string           `json:"status"`
+	Model            string           `json:"model,omitempty"`
+	IPAddress        string           `json:"ip_address,omitempty"`
+	Site             *resourceRefView `json:"site,omitempty"`
+	Building         *resourceRefView `json:"building,omitempty"`
+	Rack             *resourceRefView `json:"rack,omitempty"`
+	Zone             string           `json:"zone,omitempty"`
+}
+
+func (t *FleetTools) listMatchingMinerSnapshots(
+	ctx context.Context,
+	input resolveMinersInput,
+	filter *fleetv1.MinerListFilter,
+) ([]*fleetv1.MinerStateSnapshot, int, int32, bool, error) {
+	pageSize := input.Limit
+	if input.Query != "" {
+		pageSize = maxResolveMinersLimit
+	}
+	cursor := ""
+	scanned := 0
+	matchedScanned := 0
+	totalAvailable := int32(0)
+	totalAvailableSet := false
+	truncated := false
+	miners := make([]*fleetv1.MinerStateSnapshot, 0, min(int(input.Limit), defaultResolveMinersLimit))
+
+	for {
+		response, err := t.fleet.ListMinerStateSnapshots(ctx, connect.NewRequest(&fleetv1.ListMinerStateSnapshotsRequest{
+			PageSize: pageSize,
+			Cursor:   cursor,
+			Filter:   filter,
+		}))
+		if err != nil {
+			return nil, 0, 0, false, err
+		}
+		if !totalAvailableSet {
+			totalAvailable = response.Msg.GetTotalMiners()
+			totalAvailableSet = true
+		}
+
+		for _, miner := range response.Msg.GetMiners() {
+			scanned++
+			if !matchesResolveMinerQuery(miner, input.Query) {
+				continue
+			}
+			matchedScanned++
+			if len(miners) >= int(input.Limit) {
+				truncated = true
+				continue
+			}
+			miners = append(miners, miner)
+		}
+
+		nextCursor := response.Msg.GetCursor()
+		if nextCursor == "" {
+			break
+		}
+		if input.Query == "" || len(miners) >= int(input.Limit) || scanned >= maxResolveMinersScan {
+			truncated = true
+			break
+		}
+		cursor = nextCursor
+	}
+
+	return miners, matchedScanned, totalAvailable, truncated, nil
+}
+
+func (t *FleetTools) resolveMiners(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, filter, err := buildResolveMinersFilter(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	snapshots, matchedScanned, totalAvailable, truncated, err := t.listMatchingMinerSnapshots(ctx, input, filter)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+
+	miners := make([]resolvedMinerView, 0, len(snapshots))
+	deviceIdentifiers := make([]string, 0, len(snapshots))
+	for _, miner := range snapshots {
+		miners = append(miners, resolvedMinerViewFromSnapshot(miner))
+		deviceIdentifiers = append(deviceIdentifiers, miner.GetDeviceIdentifier())
+	}
+	payload := map[string]any{
+		"device_identifiers": deviceIdentifiers,
+		"miners":             miners,
+		"returned":           len(miners),
+		"matched_scanned":    matchedScanned,
+		"total_available":    totalAvailable,
+		"truncated":          truncated,
+	}
+	if input.Query != "" {
+		payload["query"] = input.Query
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal resolved miners: %w", err)
+	}
+	summary := fmt.Sprintf("Resolved %d miner(s)", len(miners))
+	if truncated {
+		summary += "; more matches may exist"
+	}
+	return chatdomain.ToolOutput{Content: string(content), Summary: summary}, nil
+}
+
+type actionableMinerIssueView struct {
+	DeviceIdentifier string           `json:"device_identifier"`
+	Name             string           `json:"name,omitempty"`
+	Status           string           `json:"status"`
+	Severity         string           `json:"severity"`
+	Issue            string           `json:"issue"`
+	SuggestedAction  string           `json:"suggested_action"`
+	Model            string           `json:"model,omitempty"`
+	IPAddress        string           `json:"ip_address,omitempty"`
+	Site             *resourceRefView `json:"site,omitempty"`
+	Building         *resourceRefView `json:"building,omitempty"`
+	Rack             *resourceRefView `json:"rack,omitempty"`
+	Zone             string           `json:"zone,omitempty"`
+}
+
+func (t *FleetTools) listActionableMinerIssues(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, filter, err := buildResolveMinersFilter(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	if len(filter.GetDeviceStatus()) == 0 {
+		filter.DeviceStatus = []fleetv1.DeviceStatus{
+			fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE,
+			fleetv1.DeviceStatus_DEVICE_STATUS_ERROR,
+			fleetv1.DeviceStatus_DEVICE_STATUS_NEEDS_MINING_POOL,
+			fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED,
+		}
+	}
+	snapshots, matchedScanned, totalAvailable, truncated, err := t.listMatchingMinerSnapshots(ctx, input, filter)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	issues := make([]actionableMinerIssueView, 0, len(snapshots))
+	for _, miner := range snapshots {
+		issue, include := actionableIssueFromMiner(miner)
+		if include {
+			issues = append(issues, issue)
+		}
+	}
+	payload := map[string]any{
+		"issues":          issues,
+		"returned":        len(issues),
+		"matched_scanned": matchedScanned,
+		"total_available": totalAvailable,
+		"truncated":       truncated,
+	}
+	if input.Query != "" {
+		payload["query"] = input.Query
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal actionable miner issues: %w", err)
+	}
+	summary := fmt.Sprintf("Found %d actionable miner issue(s)", len(issues))
+	if truncated {
+		summary += "; more matches may exist"
+	}
+	return chatdomain.ToolOutput{Content: string(content), Summary: summary}, nil
+}
+
+func actionableIssueFromMiner(miner *fleetv1.MinerStateSnapshot) (actionableMinerIssueView, bool) {
+	severity, issue, suggestedAction, include := actionableStatus(miner.GetDeviceStatus())
+	if !include {
+		return actionableMinerIssueView{}, false
+	}
+	base := resolvedMinerViewFromSnapshot(miner)
+	return actionableMinerIssueView{
+		DeviceIdentifier: base.DeviceIdentifier,
+		Name:             base.Name,
+		Status:           base.Status,
+		Severity:         severity,
+		Issue:            issue,
+		SuggestedAction:  suggestedAction,
+		Model:            base.Model,
+		IPAddress:        base.IPAddress,
+		Site:             base.Site,
+		Building:         base.Building,
+		Rack:             base.Rack,
+		Zone:             base.Zone,
+	}, true
+}
+
+func actionableStatus(status fleetv1.DeviceStatus) (string, string, string, bool) {
+	switch status {
+	case fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE:
+		return "high", "Miner is offline", "Check power and network reachability; if reachable, preview a reboot before executing it.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_ERROR:
+		return "high", "Miner is in an error state", "Inspect device errors and recent activity before deciding whether to reboot or repair.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_NEEDS_MINING_POOL:
+		return "medium", "Miner is missing a mining pool", "Assign an approved mining pool from the normal pool management flow.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED:
+		return "medium", "Miner reports reboot required", "Use preview_miner_action with reboot, then execute_miner_action after operator confirmation.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_UPDATING:
+		return "medium", "Miner is updating", "Monitor update progress before issuing another command.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_MAINTENANCE:
+		return "low", "Miner is in maintenance", "Confirm maintenance intent before returning it to service.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_INACTIVE:
+		return "low", "Miner is inactive or sleeping", "If this miner should be hashing, preview start_mining before executing it.", true
+	case fleetv1.DeviceStatus_DEVICE_STATUS_UNSPECIFIED, fleetv1.DeviceStatus_DEVICE_STATUS_ONLINE:
+		return "", "", "", false
+	default:
+		return "", "", "", false
+	}
+}
+
+func matchesResolveMinerQuery(miner *fleetv1.MinerStateSnapshot, query string) bool {
+	if query == "" {
+		return true
+	}
+	needle := strings.ToLower(query)
+	fields := []string{
+		miner.GetDeviceIdentifier(),
+		miner.GetName(),
+		miner.GetMacAddress(),
+		miner.GetSerialNumber(),
+		miner.GetIpAddress(),
+		miner.GetModel(),
+	}
+	for _, field := range fields {
+		if strings.Contains(strings.ToLower(field), needle) {
+			return true
+		}
+	}
+	return false
+}
+
+func resolvedMinerViewFromSnapshot(miner *fleetv1.MinerStateSnapshot) resolvedMinerView {
+	view := resolvedMinerView{
+		DeviceIdentifier: miner.GetDeviceIdentifier(),
+		Name:             miner.GetName(),
+		Status:           deviceStatusLabel(miner.GetDeviceStatus()),
+		Model:            miner.GetModel(),
+		IPAddress:        miner.GetIpAddress(),
+	}
+	if placement := miner.GetPlacement(); placement != nil {
+		view.Site = resourceRefViewFromProto(placement.GetSite())
+		view.Building = resourceRefViewFromProto(placement.GetBuilding())
+		view.Rack = resourceRefViewFromProto(placement.GetRack())
+		view.Zone = placement.GetZone()
+	}
+	return view
+}
+
+func resourceRefViewFromProto(ref *commonv1.ResourceRef) *resourceRefView {
+	if ref == nil {
+		return nil
+	}
+	return &resourceRefView{ID: ref.GetId(), Label: ref.GetLabel()}
+}
+
+func deviceStatusLabel(status fleetv1.DeviceStatus) string {
+	label := strings.TrimPrefix(status.String(), "DEVICE_STATUS_")
+	return strings.ToLower(label)
+}
+
+func rackOrderIndexLabel(orderIndex devicesetv1.RackOrderIndex) string {
+	switch orderIndex {
+	case devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_UNSPECIFIED:
+		return labelUnspecified
+	case devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_BOTTOM_LEFT:
+		return rackOrderIndexBottomLeft
+	case devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_TOP_LEFT:
+		return rackOrderIndexTopLeft
+	case devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_BOTTOM_RIGHT:
+		return rackOrderIndexBottomRight
+	case devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_TOP_RIGHT:
+		return rackOrderIndexTopRight
+	default:
+		return labelUnspecified
+	}
+}
+
+func formatRackSlotAssignments(assignments []rackSlotAssignmentInput, limit int) string {
+	parts := make([]string, 0, min(len(assignments), limit)+1)
+	for i, assignment := range assignments {
+		if i >= limit {
+			parts = append(parts, fmt.Sprintf("+%d more", len(assignments)-limit))
+			break
+		}
+		parts = append(parts, fmt.Sprintf("%s → row %d, column %d", assignment.DeviceIdentifier, assignment.Row, assignment.Column))
+	}
+	return strings.Join(parts, "; ")
+}
+
+func formatStringList(values []string, limit int) string {
+	if len(values) <= limit {
+		return strings.Join(values, ", ")
+	}
+	parts := append([]string{}, values[:limit]...)
+	parts = append(parts, fmt.Sprintf("+%d more", len(values)-limit))
+	return strings.Join(parts, ", ")
+}
+
+func minerActionTargetPhrase(selection minerActionSelectionView) string {
+	switch selection.Type {
+	case minerActionSelectorExplicit:
+		return fmt.Sprintf("%d miner(s)", len(selection.DeviceIdentifiers))
+	case minerActionSelectorAllDevices:
+		return "the whole fleet"
+	case minerActionSelectorFilter:
+		return "matching miners"
+	default:
+		return selection.Description
+	}
+}
+
+func minerActionSelectionDetail(selection minerActionSelectionView) string {
+	switch selection.Type {
+	case minerActionSelectorExplicit:
+		return formatStringList(selection.DeviceIdentifiers, 20)
+	case minerActionSelectorAllDevices:
+		return "Whole fleet command-eligible miners"
+	case minerActionSelectorFilter:
+		return formatMinerActionFilter(selection.Filter)
+	default:
+		return selection.Description
+	}
+}
+
+func formatMinerActionFilter(filter *minerActionFilterInput) string {
+	if filter == nil {
+		return "All command-eligible miners matching an empty filter"
+	}
+	parts := make([]string, 0, 8)
+	if len(filter.DeviceStatuses) > 0 {
+		parts = append(parts, "statuses: "+strings.Join(filter.DeviceStatuses, ", "))
+	}
+	if len(filter.SiteIDs) > 0 {
+		parts = append(parts, "site IDs: "+formatInt64List(filter.SiteIDs, 20))
+	}
+	if filter.IncludeUnassigned {
+		parts = append(parts, "include unassigned")
+	}
+	if len(filter.BuildingIDs) > 0 {
+		parts = append(parts, "building IDs: "+formatInt64List(filter.BuildingIDs, 20))
+	}
+	if filter.IncludeNoBuilding {
+		parts = append(parts, "include no building")
+	}
+	if len(filter.RackIDs) > 0 {
+		parts = append(parts, "rack IDs: "+formatInt64List(filter.RackIDs, 20))
+	}
+	if filter.IncludeNoRack {
+		parts = append(parts, "include no rack")
+	}
+	if len(filter.Models) > 0 {
+		parts = append(parts, "models: "+formatStringList(filter.Models, 20))
+	}
+	if len(filter.IPCIDRs) > 0 {
+		parts = append(parts, "IP CIDRs: "+formatStringList(filter.IPCIDRs, 20))
+	}
+	if len(parts) == 0 {
+		return "All command-eligible miners matching an empty filter"
+	}
+	return strings.Join(parts, "; ")
+}
+
+func formatInt64List(values []int64, limit int) string {
+	parts := make([]string, 0, min(len(values), limit)+1)
+	for i, value := range values {
+		if i >= limit {
+			parts = append(parts, fmt.Sprintf("+%d more", len(values)-limit))
+			break
+		}
+		parts = append(parts, fmt.Sprint(value))
+	}
+	return strings.Join(parts, ", ")
+}
+
+type siteInventoryView struct {
+	ID             int64  `json:"id"`
+	Name           string `json:"name"`
+	DeviceCount    int64  `json:"device_count"`
+	BuildingCount  int64  `json:"building_count"`
+	RackCount      int64  `json:"rack_count"`
+	Infrastructure int64  `json:"infrastructure_device_count"`
+}
+
+func (t *FleetTools) siteInventory(ctx context.Context, siteID *int64) ([]siteInventoryView, error) {
+	response, err := t.sites.ListSites(ctx, connect.NewRequest(&sitesv1.ListSitesRequest{}))
+	if err != nil {
+		return nil, err
+	}
+	sites := make([]siteInventoryView, 0, len(response.Msg.GetSites()))
+	for _, item := range response.Msg.GetSites() {
+		if siteID != nil && item.GetSite().GetId() != *siteID {
+			continue
+		}
+		sites = append(sites, siteInventoryView{
+			ID:             item.GetSite().GetId(),
+			Name:           item.GetSite().GetName(),
+			DeviceCount:    item.GetDeviceCount(),
+			BuildingCount:  item.GetBuildingCount(),
+			RackCount:      item.GetRackCount(),
+			Infrastructure: item.GetInfrastructureDeviceCount(),
+		})
+	}
+	return sites, nil
+}
+
+func (t *FleetTools) listSites(ctx context.Context) (chatdomain.ToolOutput, error) {
+	sites, err := t.siteInventory(ctx, nil)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	content, err := json.Marshal(map[string]any{"sites": sites})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal sites: %w", err)
+	}
+	return chatdomain.ToolOutput{Content: string(content), Summary: fmt.Sprintf("Read %d sites", len(sites))}, nil
+}
+
+func (t *FleetTools) listPools(ctx context.Context) (chatdomain.ToolOutput, error) {
+	response, err := t.pools.ListPools(ctx, connect.NewRequest(&poolsv1.ListPoolsRequest{}))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	type poolView struct {
+		Name string `json:"name"`
+	}
+	pools := make([]poolView, 0, len(response.Msg.GetPools()))
+	for _, pool := range response.Msg.GetPools() {
+		pools = append(pools, poolView{Name: pool.GetPoolName()})
+	}
+	content, err := json.Marshal(map[string]any{"pools": pools})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal pools: %w", err)
+	}
+	return chatdomain.ToolOutput{Content: string(content), Summary: fmt.Sprintf("Read %d pools", len(pools))}, nil
+}
+
+func (t *FleetTools) listRacks(ctx context.Context) (chatdomain.ToolOutput, error) {
+	response, err := t.deviceSets.ListDeviceSets(ctx, connect.NewRequest(&devicesetv1.ListDeviceSetsRequest{
+		Type:     devicesetv1.DeviceSetType_DEVICE_SET_TYPE_RACK,
+		PageSize: 1000,
+	}))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	type rackView struct {
+		ID              int64  `json:"id"`
+		Label           string `json:"label"`
+		Rows            int32  `json:"rows"`
+		Columns         int32  `json:"columns"`
+		NumberingOrigin string `json:"numbering_origin"`
+		DeviceCount     int32  `json:"device_count"`
+		Site            string `json:"site,omitempty"`
+		Building        string `json:"building,omitempty"`
+		Zone            string `json:"zone,omitempty"`
+	}
+	racks := make([]rackView, 0, len(response.Msg.GetDeviceSets()))
+	for _, rack := range response.Msg.GetDeviceSets() {
+		rackInfo := rack.GetRackInfo()
+		view := rackView{
+			ID:              rack.GetId(),
+			Label:           rack.GetLabel(),
+			Rows:            rackInfo.GetRows(),
+			Columns:         rackInfo.GetColumns(),
+			NumberingOrigin: rackOrderIndexLabel(rackInfo.GetOrderIndex()),
+			DeviceCount:     rack.GetDeviceCount(),
+			Zone:            rackInfo.GetZone(),
+		}
+		if placement := rack.GetPlacement(); placement != nil {
+			view.Site = placement.GetSite().GetLabel()
+			view.Building = placement.GetBuilding().GetLabel()
+		}
+		racks = append(racks, view)
+	}
+	content, err := json.Marshal(map[string]any{"racks": racks})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal racks: %w", err)
+	}
+	return chatdomain.ToolOutput{Content: string(content), Summary: fmt.Sprintf("Read %d racks", len(racks))}, nil
+}
+
+type rackSlotView struct {
+	DeviceIdentifier string `json:"device_identifier"`
+	Row              int32  `json:"row"`
+	Column           int32  `json:"column"`
+}
+
+func rackSlotViewFromProto(slot *devicesetv1.RackSlot) rackSlotView {
+	return rackSlotView{
+		DeviceIdentifier: slot.GetDeviceIdentifier(),
+		Row:              slot.GetPosition().GetRow(),
+		Column:           slot.GetPosition().GetColumn(),
+	}
+}
+
+func rackSlotAssignmentView(assignment rackSlotAssignmentInput) rackSlotView {
+	return rackSlotView{
+		DeviceIdentifier: assignment.DeviceIdentifier,
+		Row:              assignment.Row,
+		Column:           assignment.Column,
+	}
+}
+
+func (t *FleetTools) getRackSlots(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	rackID, request, err := buildGetRackSlotsRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.deviceSets.GetRackSlots(ctx, connect.NewRequest(request))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	slots := make([]rackSlotView, 0, len(response.Msg.GetSlots()))
+	for _, slot := range response.Msg.GetSlots() {
+		slots = append(slots, rackSlotViewFromProto(slot))
+	}
+	content, err := json.Marshal(map[string]any{
+		"rack_id":        rackID,
+		"occupied_count": len(slots),
+		"occupied_slots": slots,
+	})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal rack slots: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Read %d occupied slot(s) for rack %d", len(slots), rackID),
+	}, nil
+}
+
+type rackSlotPositionView struct {
+	Row    int32 `json:"row"`
+	Column int32 `json:"column"`
+}
+
+type rackHealthMinerView struct {
+	resolvedMinerView
+	Slot *rackSlotPositionView `json:"slot,omitempty"`
+}
+
+func (t *FleetTools) getRackHealth(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, err := buildRackHealthInput(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	rack, members, slots, err := t.loadRackSlotState(ctx, input.RackID)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	resolveInput := resolveMinersInput{
+		RackIDs: []int64{input.RackID},
+		Limit:   input.Limit,
+	}
+	filter := &fleetv1.MinerListFilter{RackIds: []int64{input.RackID}}
+	snapshots, _, totalAvailable, truncated, err := t.listMatchingMinerSnapshots(ctx, resolveInput, filter)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	slotsByMiner := make(map[string]rackSlotPositionView, len(slots))
+	for _, slot := range slots {
+		slotsByMiner[slot.GetDeviceIdentifier()] = rackSlotPositionView{
+			Row:    slot.GetPosition().GetRow(),
+			Column: slot.GetPosition().GetColumn(),
+		}
+	}
+	statusCounts := emptyHealthStatusCounts()
+	miners := make([]rackHealthMinerView, 0, len(snapshots))
+	for _, miner := range snapshots {
+		statusCounts[healthStatusKey(miner.GetDeviceStatus())]++
+		view := rackHealthMinerView{resolvedMinerView: resolvedMinerViewFromSnapshot(miner)}
+		if slot, ok := slotsByMiner[miner.GetDeviceIdentifier()]; ok {
+			view.Slot = &slot
+		}
+		miners = append(miners, view)
+	}
+	rackInfo := rack.GetRackInfo()
+	payload := map[string]any{
+		"rack": map[string]any{
+			"id":               rack.GetId(),
+			"label":            rack.GetLabel(),
+			"rows":             rackInfo.GetRows(),
+			"columns":          rackInfo.GetColumns(),
+			"numbering_origin": rackOrderIndexLabel(rackInfo.GetOrderIndex()),
+			"zone":             rackInfo.GetZone(),
+			"device_count":     rack.GetDeviceCount(),
+		},
+		"member_count":        len(members),
+		"occupied_slot_count": len(slots),
+		"status_counts":       statusCounts,
+		"miners":              miners,
+		"returned":            len(miners),
+		"total_available":     totalAvailable,
+		"truncated":           truncated,
+	}
+	if placement := rack.GetPlacement(); placement != nil {
+		payload["placement"] = map[string]any{
+			"site":     resourceRefViewFromProto(placement.GetSite()),
+			"building": resourceRefViewFromProto(placement.GetBuilding()),
+		}
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal rack health: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Read health for rack %d with %d member(s)", input.RackID, len(members)),
+	}, nil
+}
+
+func emptyHealthStatusCounts() map[string]int {
+	return map[string]int{
+		"hashing":            0,
+		"broken":             0,
+		"offline":            0,
+		"sleeping":           0,
+		"maintenance":        0,
+		"needs_mining_pool":  0,
+		"updating":           0,
+		"reboot_required":    0,
+		"status_unspecified": 0,
+	}
+}
+
+func healthStatusKey(status fleetv1.DeviceStatus) string {
+	switch status {
+	case fleetv1.DeviceStatus_DEVICE_STATUS_ONLINE:
+		return "hashing"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_ERROR:
+		return "broken"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_INACTIVE:
+		return "sleeping"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE:
+		return "offline"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_MAINTENANCE:
+		return "maintenance"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_NEEDS_MINING_POOL:
+		return "needs_mining_pool"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_UPDATING:
+		return "updating"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED:
+		return "reboot_required"
+	case fleetv1.DeviceStatus_DEVICE_STATUS_UNSPECIFIED:
+		return "status_unspecified"
+	default:
+		return "status_unspecified"
+	}
+}
+
+func (t *FleetTools) setRackSlots(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, err := buildSetRackSlotsInput(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	rack, members, existingSlots, err := t.loadRackSlotState(ctx, input.RackID)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	if err := validateSetRackSlots(input, rack, members, existingSlots); err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+
+	requestedDevices := make(map[string]struct{}, len(input.SlotAssignments))
+	for _, assignment := range input.SlotAssignments {
+		requestedDevices[assignment.DeviceIdentifier] = struct{}{}
+	}
+	assignmentViews := make([]rackSlotView, 0, len(input.SlotAssignments))
+	for _, assignment := range input.SlotAssignments {
+		assignmentViews = append(assignmentViews, rackSlotAssignmentView(assignment))
+	}
+
+	for identifier := range requestedDevices {
+		if _, err := t.deviceSets.ClearRackSlotPosition(ctx, connect.NewRequest(&devicesetv1.ClearRackSlotPositionRequest{
+			DeviceSetId:      input.RackID,
+			DeviceIdentifier: identifier,
+		})); err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+	}
+	for _, assignment := range input.SlotAssignments {
+		if _, err := t.deviceSets.SetRackSlotPosition(ctx, connect.NewRequest(&devicesetv1.SetRackSlotPositionRequest{
+			DeviceSetId:      input.RackID,
+			DeviceIdentifier: assignment.DeviceIdentifier,
+			Position:         &devicesetv1.RackSlotPosition{Row: assignment.Row, Column: assignment.Column},
+		})); err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+	}
+	content, err := json.Marshal(map[string]any{
+		"applied":          true,
+		"rack_id":          input.RackID,
+		"rack_label":       rack.GetLabel(),
+		"assigned_count":   len(input.SlotAssignments),
+		"slot_assignments": assignmentViews,
+	})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal assigned rack slots: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Assigned %d slot(s) in rack %d", len(input.SlotAssignments), input.RackID),
+	}, nil
+}
+
+func (t *FleetTools) clearRackSlots(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, err := buildClearRackSlotsInput(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	rack, members, existingSlots, err := t.loadRackSlotState(ctx, input.RackID)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	memberSet := makeStringSet(members)
+	for _, identifier := range input.DeviceIdentifiers {
+		if _, ok := memberSet[identifier]; !ok {
+			return chatdomain.ToolOutput{}, fmt.Errorf("invalid clear_rack_slots arguments: miner %q is not assigned to rack %d", identifier, input.RackID)
+		}
+	}
+	clearDevices := makeStringSet(input.DeviceIdentifiers)
+	cleared := 0
+	for _, slot := range existingSlots {
+		if _, shouldClear := clearDevices[slot.GetDeviceIdentifier()]; shouldClear {
+			cleared++
+		}
+	}
+	for _, identifier := range input.DeviceIdentifiers {
+		if _, err := t.deviceSets.ClearRackSlotPosition(ctx, connect.NewRequest(&devicesetv1.ClearRackSlotPositionRequest{
+			DeviceSetId:      input.RackID,
+			DeviceIdentifier: identifier,
+		})); err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+	}
+	content, err := json.Marshal(map[string]any{
+		"cleared":            true,
+		"rack_id":            input.RackID,
+		"rack_label":         rack.GetLabel(),
+		"requested_count":    len(input.DeviceIdentifiers),
+		"cleared_count":      cleared,
+		"device_identifiers": input.DeviceIdentifiers,
+	})
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal cleared rack slots: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Cleared %d slot(s) in rack %d", cleared, input.RackID),
+	}, nil
+}
+
+func (t *FleetTools) loadRackSlotState(ctx context.Context, rackID int64) (*devicesetv1.DeviceSet, []string, []*devicesetv1.RackSlot, error) {
+	rackResponse, err := t.deviceSets.GetDeviceSet(ctx, connect.NewRequest(&devicesetv1.GetDeviceSetRequest{DeviceSetId: rackID}))
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	rack := rackResponse.Msg.GetDeviceSet()
+	if rack.GetType() != devicesetv1.DeviceSetType_DEVICE_SET_TYPE_RACK {
+		return nil, nil, nil, fmt.Errorf("invalid rack slot arguments: device set %d is not a rack", rackID)
+	}
+
+	members, err := t.listRackMemberIdentifiers(ctx, rackID)
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	slotsResponse, err := t.deviceSets.GetRackSlots(ctx, connect.NewRequest(&devicesetv1.GetRackSlotsRequest{DeviceSetId: rackID}))
+	if err != nil {
+		return nil, nil, nil, err
+	}
+	return rack, members, slotsResponse.Msg.GetSlots(), nil
+}
+
+func (t *FleetTools) listRackMemberIdentifiers(ctx context.Context, rackID int64) ([]string, error) {
+	const pageSize = 1000
+	members := make([]string, 0)
+	pageToken := ""
+	for {
+		response, err := t.deviceSets.ListDeviceSetMembers(ctx, connect.NewRequest(&devicesetv1.ListDeviceSetMembersRequest{
+			DeviceSetId: rackID,
+			PageSize:    pageSize,
+			PageToken:   pageToken,
+		}))
+		if err != nil {
+			return nil, err
+		}
+		for _, member := range response.Msg.GetMembers() {
+			members = append(members, member.GetDeviceIdentifier())
+		}
+		pageToken = response.Msg.GetNextPageToken()
+		if pageToken == "" {
+			break
+		}
+	}
+	return members, nil
+}
+
+func validateSetRackSlots(input setRackSlotsInput, rack *devicesetv1.DeviceSet, members []string, existingSlots []*devicesetv1.RackSlot) error {
+	rackInfo := rack.GetRackInfo()
+	if rackInfo.GetRows() <= 0 || rackInfo.GetColumns() <= 0 {
+		return fmt.Errorf("invalid set_rack_slots arguments: rack %d does not have a usable slot layout", input.RackID)
+	}
+	memberSet := makeStringSet(members)
+	requestedDevices := make(map[string]struct{}, len(input.SlotAssignments))
+	requestedPositions := make(map[string]rackSlotAssignmentInput, len(input.SlotAssignments))
+	for _, assignment := range input.SlotAssignments {
+		if _, ok := memberSet[assignment.DeviceIdentifier]; !ok {
+			return fmt.Errorf("invalid set_rack_slots arguments: miner %q is not assigned to rack %d", assignment.DeviceIdentifier, input.RackID)
+		}
+		if assignment.Row >= rackInfo.GetRows() {
+			return fmt.Errorf("invalid set_rack_slots arguments: row %d is out of bounds for rack %d (%d rows, 0-indexed)", assignment.Row, input.RackID, rackInfo.GetRows())
+		}
+		if assignment.Column >= rackInfo.GetColumns() {
+			return fmt.Errorf("invalid set_rack_slots arguments: column %d is out of bounds for rack %d (%d columns, 0-indexed)", assignment.Column, input.RackID, rackInfo.GetColumns())
+		}
+		requestedDevices[assignment.DeviceIdentifier] = struct{}{}
+		requestedPositions[rackSlotPositionKey(assignment.Row, assignment.Column)] = assignment
+	}
+	for _, slot := range existingSlots {
+		assignment, positionRequested := requestedPositions[rackSlotPositionKey(slot.GetPosition().GetRow(), slot.GetPosition().GetColumn())]
+		if !positionRequested {
+			continue
+		}
+		if slot.GetDeviceIdentifier() == assignment.DeviceIdentifier {
+			continue
+		}
+		if _, movingOccupant := requestedDevices[slot.GetDeviceIdentifier()]; movingOccupant {
+			continue
+		}
+		return fmt.Errorf(
+			"invalid set_rack_slots arguments: slot (%d,%d) is already occupied by miner %q; include that miner in the request or choose an empty slot",
+			assignment.Row,
+			assignment.Column,
+			slot.GetDeviceIdentifier(),
+		)
+	}
+	return nil
+}
+
+func makeStringSet(values []string) map[string]struct{} {
+	set := make(map[string]struct{}, len(values))
+	for _, value := range values {
+		set[value] = struct{}{}
+	}
+	return set
+}
+
+func rackSlotPositionKey(row, column int32) string {
+	return fmt.Sprintf("%d:%d", row, column)
+}
+
+type unsupportedMinerGroupView struct {
+	FirmwareVersion string `json:"firmware_version,omitempty"`
+	Model           string `json:"model,omitempty"`
+	Count           int32  `json:"count"`
+}
+
+func (t *FleetTools) previewMinerAction(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	if t.commands == nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("preview_miner_action is unavailable")
+	}
+	input, commandType, selector, selection, err := buildMinerActionInput(arguments, "preview_miner_action")
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.commands.CheckCommandCapabilities(ctx, connect.NewRequest(&minercommandv1.CheckCommandCapabilitiesRequest{
+		CommandType:    commandType,
+		DeviceSelector: selector,
+	}))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	unsupportedGroups := make([]unsupportedMinerGroupView, 0, len(response.Msg.GetUnsupportedGroups()))
+	for _, group := range response.Msg.GetUnsupportedGroups() {
+		unsupportedGroups = append(unsupportedGroups, unsupportedMinerGroupView{
+			FirmwareVersion: group.GetFirmwareVersion(),
+			Model:           group.GetModel(),
+			Count:           group.GetCount(),
+		})
+	}
+	payload := map[string]any{
+		"action":                       normalizeToolEnum(input.Action),
+		"selector":                     selection,
+		"total_count":                  response.Msg.GetTotalCount(),
+		"supported_count":              response.Msg.GetSupportedCount(),
+		"unsupported_count":            response.Msg.GetUnsupportedCount(),
+		"all_supported":                response.Msg.GetAllSupported(),
+		"none_supported":               response.Msg.GetNoneSupported(),
+		"supported_device_identifiers": response.Msg.GetSupportedDeviceIdentifiers(),
+		"unsupported_groups":           unsupportedGroups,
+	}
+	if selection.Type == minerActionSelectorExplicit {
+		payload["requested_device_identifiers"] = selection.DeviceIdentifiers
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal miner action preview: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Previewed %s for %s", strings.ReplaceAll(normalizeToolEnum(input.Action), "_", " "), minerActionTargetPhrase(selection)),
+	}, nil
+}
+
+func (t *FleetTools) executeMinerAction(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	if t.commands == nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("execute_miner_action is unavailable")
+	}
+	input, commandType, selector, selection, err := buildMinerActionInput(arguments, "execute_miner_action")
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	var batchIdentifier string
+	switch commandType {
+	case minercommandv1.CommandType_COMMAND_TYPE_REBOOT:
+		response, err := t.commands.Reboot(ctx, connect.NewRequest(&minercommandv1.RebootRequest{DeviceSelector: selector}))
+		if err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+		batchIdentifier = response.Msg.GetBatchIdentifier()
+	case minercommandv1.CommandType_COMMAND_TYPE_START_MINING:
+		response, err := t.commands.StartMining(ctx, connect.NewRequest(&minercommandv1.StartMiningRequest{DeviceSelector: selector}))
+		if err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+		batchIdentifier = response.Msg.GetBatchIdentifier()
+	case minercommandv1.CommandType_COMMAND_TYPE_STOP_MINING:
+		response, err := t.commands.StopMining(ctx, connect.NewRequest(&minercommandv1.StopMiningRequest{DeviceSelector: selector}))
+		if err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+		batchIdentifier = response.Msg.GetBatchIdentifier()
+	case minercommandv1.CommandType_COMMAND_TYPE_BLINK_LED:
+		response, err := t.commands.BlinkLED(ctx, connect.NewRequest(&minercommandv1.BlinkLEDRequest{DeviceSelector: selector}))
+		if err != nil {
+			return chatdomain.ToolOutput{}, err
+		}
+		batchIdentifier = response.Msg.GetBatchIdentifier()
+	case minercommandv1.CommandType_COMMAND_TYPE_UNSPECIFIED,
+		minercommandv1.CommandType_COMMAND_TYPE_SET_COOLING_MODE,
+		minercommandv1.CommandType_COMMAND_TYPE_UPDATE_MINING_POOLS,
+		minercommandv1.CommandType_COMMAND_TYPE_DOWNLOAD_LOGS,
+		minercommandv1.CommandType_COMMAND_TYPE_FIRMWARE_UPDATE,
+		minercommandv1.CommandType_COMMAND_TYPE_SET_POWER_TARGET,
+		minercommandv1.CommandType_COMMAND_TYPE_UPDATE_MINER_PASSWORD,
+		minercommandv1.CommandType_COMMAND_TYPE_CURTAIL,
+		minercommandv1.CommandType_COMMAND_TYPE_UNCURTAIL:
+		return chatdomain.ToolOutput{}, fmt.Errorf("invalid execute_miner_action arguments: unsupported action %q", input.Action)
+	default:
+		return chatdomain.ToolOutput{}, fmt.Errorf("invalid execute_miner_action arguments: unsupported action %q", input.Action)
+	}
+	action := normalizeToolEnum(input.Action)
+	payload := map[string]any{
+		"executed":                   true,
+		"action":                     action,
+		"selector":                   selection,
+		"command_batch_identifier":   batchIdentifier,
+		"command_status_lookup_hint": "Use the existing command activity UI to inspect per-miner command results.",
+	}
+	if selection.Type == minerActionSelectorExplicit {
+		payload["selected_count"] = len(selection.DeviceIdentifiers)
+		payload["device_identifiers"] = selection.DeviceIdentifiers
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal executed miner action: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Submitted %s for %s", strings.ReplaceAll(action, "_", " "), minerActionTargetPhrase(selection)),
+	}, nil
+}
+
+func (t *FleetTools) previewDowntimeWindow(_ context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, request, err := buildDowntimeWindowRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	payload := downtimeWindowPayload(input, request, 0)
+	payload["will_create"] = false
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal downtime window preview: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Previewed downtime schedule %q", input.Name),
+	}, nil
+}
+
+func (t *FleetTools) createDowntimeWindow(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	if t.schedules == nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("create_downtime_window is unavailable")
+	}
+	input, request, err := buildDowntimeWindowRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.schedules.CreateSchedule(ctx, connect.NewRequest(request))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	payload := downtimeWindowPayload(input, request, response.Msg.GetSchedule().GetId())
+	payload["created"] = true
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal created downtime window: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Created schedule %q", input.Name),
+	}, nil
+}
+
+func downtimeWindowPayload(input downtimeWindowInput, request *schedulev1.CreateScheduleRequest, scheduleID int64) map[string]any {
+	payload := map[string]any{
+		"schedule_id":   scheduleID,
+		"name":          request.GetName(),
+		"action":        scheduleActionLabel(request.GetAction()),
+		"schedule_type": "one_time",
+		"start_date":    request.GetStartDate(),
+		"start_time":    request.GetStartTime(),
+		"end_date":      request.GetEndDate(),
+		"end_time":      request.GetEndTime(),
+		"timezone":      request.GetTimezone(),
+		"targets":       input.Targets,
+		"target_count":  len(request.GetTargets()),
+		"operator_note": downtimeActionNote(request),
+	}
+	if request.GetAction() == schedulev1.ScheduleAction_SCHEDULE_ACTION_SET_POWER_TARGET {
+		payload["power_target_mode"] = powerTargetModeLabel(request.GetActionConfig().GetMode())
+	}
+	return payload
+}
+
+func downtimeActionNote(request *schedulev1.CreateScheduleRequest) string {
+	switch request.GetAction() {
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_SLEEP:
+		return "Sleep schedules stop mining at start_time and do not automatically resume at end_time."
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_SET_POWER_TARGET:
+		if request.GetEndTime() != "" {
+			return "Set-power-target schedules with end_time revert through the existing scheduler behavior."
+		}
+		return "Set-power-target schedules without end_time do not define a bounded revert window."
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_REBOOT:
+		return "Reboot schedules run once at start_time."
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_UNSPECIFIED:
+		return "Schedule behavior follows the existing schedule service."
+	default:
+		return "Schedule behavior follows the existing schedule service."
+	}
+}
+
+func scheduleActionLabel(action schedulev1.ScheduleAction) string {
+	switch action {
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_SET_POWER_TARGET:
+		return "set_power_target"
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_REBOOT:
+		return "reboot"
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_SLEEP:
+		return "sleep"
+	case schedulev1.ScheduleAction_SCHEDULE_ACTION_UNSPECIFIED:
+		return labelUnspecified
+	default:
+		return labelUnspecified
+	}
+}
+
+func powerTargetModeLabel(mode schedulev1.PowerTargetMode) string {
+	switch mode {
+	case schedulev1.PowerTargetMode_POWER_TARGET_MODE_DEFAULT:
+		return "default"
+	case schedulev1.PowerTargetMode_POWER_TARGET_MODE_MAX:
+		return "max"
+	case schedulev1.PowerTargetMode_POWER_TARGET_MODE_UNSPECIFIED:
+		return labelUnspecified
+	default:
+		return labelUnspecified
+	}
+}
+
+func formatDowntimeTargets(targets []downtimeTargetInput, limit int) string {
+	parts := make([]string, 0, min(len(targets), limit)+1)
+	for i, target := range targets {
+		if i >= limit {
+			parts = append(parts, fmt.Sprintf("+%d more", len(targets)-limit))
+			break
+		}
+		parts = append(parts, fmt.Sprintf("%s:%s", normalizeToolEnum(target.Type), target.TargetID))
+	}
+	return strings.Join(parts, ", ")
+}
+
+func (t *FleetTools) createSite(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, request, err := buildCreateSiteRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.sites.CreateSite(ctx, connect.NewRequest(request))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	warnings := response.Msg.GetNetworkConfigWarnings()
+	if warnings == nil {
+		warnings = []string{}
+	}
+	payload := map[string]any{
+		"created":  true,
+		"site_id":  response.Msg.GetSite().GetId(),
+		"name":     response.Msg.GetSite().GetName(),
+		"warnings": warnings,
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal created site: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Created site %q", input.Name),
+	}, nil
+}
+
+func (t *FleetTools) createRack(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, request, err := buildCreateRackRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.deviceSets.CreateDeviceSet(ctx, connect.NewRequest(request))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	payload := map[string]any{
+		"created": true,
+		"rack_id": response.Msg.GetDeviceSet().GetId(),
+		"label":   response.Msg.GetDeviceSet().GetLabel(),
+		"rows":    input.Rows,
+		"columns": input.Columns,
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal created rack: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Created rack %q", input.Label),
+	}, nil
+}
+
+func (t *FleetTools) moveMinersToRack(ctx context.Context, arguments json.RawMessage) (chatdomain.ToolOutput, error) {
+	input, request, err := buildMoveMinersRequest(arguments)
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	response, err := t.deviceSets.AssignDevicesToRack(ctx, connect.NewRequest(request))
+	if err != nil {
+		return chatdomain.ToolOutput{}, err
+	}
+	if len(response.Msg.GetConflicts()) > 0 {
+		conflictingMiners := make([]string, 0, len(response.Msg.GetConflicts()))
+		for _, conflict := range response.Msg.GetConflicts() {
+			conflictingMiners = append(conflictingMiners, conflict.GetDeviceIdentifier())
+		}
+		content, marshalErr := json.Marshal(map[string]any{
+			"moved":              false,
+			"requires_force":     true,
+			"conflicting_miners": conflictingMiners,
+			"reason":             "destination rack is unassigned; moving these miners would clear their site and building",
+		})
+		if marshalErr != nil {
+			return chatdomain.ToolOutput{}, fmt.Errorf("marshal rack conflicts: %w", marshalErr)
+		}
+		return chatdomain.ToolOutput{
+			Content: string(content),
+			Summary: fmt.Sprintf("No miners moved; %d placement conflict(s) need confirmation", len(conflictingMiners)),
+		}, nil
+	}
+	payload := map[string]any{
+		"moved":                 true,
+		"target_rack_id":        input.TargetRackID,
+		"assigned_count":        response.Msg.GetAssignedCount(),
+		"removed_count":         response.Msg.GetRemovedCount(),
+		"site_reassigned_count": response.Msg.GetSiteReassignedCount(),
+	}
+	content, err := json.Marshal(payload)
+	if err != nil {
+		return chatdomain.ToolOutput{}, fmt.Errorf("marshal moved miners: %w", err)
+	}
+	return chatdomain.ToolOutput{
+		Content: string(content),
+		Summary: fmt.Sprintf("Moved %d miner(s) to rack %d", response.Msg.GetAssignedCount(), input.TargetRackID),
+	}, nil
+}
diff --git a/server/internal/handlers/chat/tools_test.go b/server/internal/handlers/chat/tools_test.go
new file mode 100644
index 000000000..c0f7a01ed
--- /dev/null
+++ b/server/internal/handlers/chat/tools_test.go
@@ -0,0 +1,880 @@
+package chat
+
+import (
+	"context"
+	"encoding/json"
+	"testing"
+
+	"connectrpc.com/connect"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	commonv1 "github.com/block/proto-fleet/server/generated/grpc/common/v1"
+	devicesetv1 "github.com/block/proto-fleet/server/generated/grpc/device_set/v1"
+	fleetv1 "github.com/block/proto-fleet/server/generated/grpc/fleetmanagement/v1"
+	minercommandv1 "github.com/block/proto-fleet/server/generated/grpc/minercommand/v1"
+	poolsv1 "github.com/block/proto-fleet/server/generated/grpc/pools/v1"
+	schedulev1 "github.com/block/proto-fleet/server/generated/grpc/schedule/v1"
+	sitesv1 "github.com/block/proto-fleet/server/generated/grpc/sites/v1"
+	telemetryv1 "github.com/block/proto-fleet/server/generated/grpc/telemetry/v1"
+	chatdomain "github.com/block/proto-fleet/server/internal/domain/chat"
+)
+
+type staticFleetHandler struct {
+	miners        []*fleetv1.MinerStateSnapshot
+	totalMiners   int32
+	cursor        string
+	counts        *fleetv1.GetMinerStateCountsResponse
+	countRequests []*fleetv1.GetMinerStateCountsRequest
+	listRequests  []*fleetv1.ListMinerStateSnapshotsRequest
+}
+
+func (h *staticFleetHandler) GetMinerStateCounts(_ context.Context, request *connect.Request[fleetv1.GetMinerStateCountsRequest]) (*connect.Response[fleetv1.GetMinerStateCountsResponse], error) {
+	h.countRequests = append(h.countRequests, request.Msg)
+	if h.counts != nil {
+		return connect.NewResponse(h.counts), nil
+	}
+	return connect.NewResponse(&fleetv1.GetMinerStateCountsResponse{}), nil
+}
+
+func (h *staticFleetHandler) ListMinerStateSnapshots(_ context.Context, request *connect.Request[fleetv1.ListMinerStateSnapshotsRequest]) (*connect.Response[fleetv1.ListMinerStateSnapshotsResponse], error) {
+	h.listRequests = append(h.listRequests, request.Msg)
+	return connect.NewResponse(&fleetv1.ListMinerStateSnapshotsResponse{
+		Miners:      h.miners,
+		TotalMiners: h.totalMiners,
+		Cursor:      h.cursor,
+	}), nil
+}
+
+type staticPoolsHandler struct {
+	pools []*poolsv1.Pool
+}
+
+type recordingSitesHandler struct {
+	createRequest *sitesv1.CreateSiteRequest
+	sites         []*sitesv1.SiteWithCounts
+}
+
+func (h *recordingSitesHandler) ListSites(context.Context, *connect.Request[sitesv1.ListSitesRequest]) (*connect.Response[sitesv1.ListSitesResponse], error) {
+	return connect.NewResponse(&sitesv1.ListSitesResponse{Sites: h.sites}), nil
+}
+
+func (h *recordingSitesHandler) CreateSite(_ context.Context, request *connect.Request[sitesv1.CreateSiteRequest]) (*connect.Response[sitesv1.CreateSiteResponse], error) {
+	h.createRequest = request.Msg
+	return connect.NewResponse(&sitesv1.CreateSiteResponse{Site: &sitesv1.Site{Id: 12, Name: request.Msg.GetName()}}), nil
+}
+
+type recordingDeviceSetsHandler struct {
+	createRequest *devicesetv1.CreateDeviceSetRequest
+	moveRequest   *devicesetv1.AssignDevicesToRackRequest
+	setRequests   []*devicesetv1.SetRackSlotPositionRequest
+	clearRequests []*devicesetv1.ClearRackSlotPositionRequest
+	rack          *devicesetv1.DeviceSet
+	members       []*devicesetv1.DeviceSetMember
+	slots         []*devicesetv1.RackSlot
+}
+
+func (h *recordingDeviceSetsHandler) GetDeviceSet(context.Context, *connect.Request[devicesetv1.GetDeviceSetRequest]) (*connect.Response[devicesetv1.GetDeviceSetResponse], error) {
+	rack := h.rack
+	if rack == nil {
+		rack = &devicesetv1.DeviceSet{
+			Id:    21,
+			Type:  devicesetv1.DeviceSetType_DEVICE_SET_TYPE_RACK,
+			Label: "A1",
+			TypeDetails: &devicesetv1.DeviceSet_RackInfo{RackInfo: &devicesetv1.RackInfo{
+				Rows:        4,
+				Columns:     6,
+				OrderIndex:  devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_TOP_LEFT,
+				CoolingType: devicesetv1.RackCoolingType_RACK_COOLING_TYPE_AIR,
+			}},
+		}
+	}
+	return connect.NewResponse(&devicesetv1.GetDeviceSetResponse{DeviceSet: rack}), nil
+}
+
+func (*recordingDeviceSetsHandler) ListDeviceSets(context.Context, *connect.Request[devicesetv1.ListDeviceSetsRequest]) (*connect.Response[devicesetv1.ListDeviceSetsResponse], error) {
+	return connect.NewResponse(&devicesetv1.ListDeviceSetsResponse{}), nil
+}
+
+func (h *recordingDeviceSetsHandler) ListDeviceSetMembers(context.Context, *connect.Request[devicesetv1.ListDeviceSetMembersRequest]) (*connect.Response[devicesetv1.ListDeviceSetMembersResponse], error) {
+	members := h.members
+	if members == nil {
+		members = []*devicesetv1.DeviceSetMember{
+			{DeviceIdentifier: "miner-a"},
+			{DeviceIdentifier: "miner-b"},
+		}
+	}
+	return connect.NewResponse(&devicesetv1.ListDeviceSetMembersResponse{Members: members}), nil
+}
+
+func (h *recordingDeviceSetsHandler) CreateDeviceSet(_ context.Context, request *connect.Request[devicesetv1.CreateDeviceSetRequest]) (*connect.Response[devicesetv1.CreateDeviceSetResponse], error) {
+	h.createRequest = request.Msg
+	return connect.NewResponse(&devicesetv1.CreateDeviceSetResponse{DeviceSet: &devicesetv1.DeviceSet{Id: 21, Label: request.Msg.GetLabel()}}), nil
+}
+
+func (h *recordingDeviceSetsHandler) AssignDevicesToRack(_ context.Context, request *connect.Request[devicesetv1.AssignDevicesToRackRequest]) (*connect.Response[devicesetv1.AssignDevicesToRackResponse], error) {
+	h.moveRequest = request.Msg
+	return connect.NewResponse(&devicesetv1.AssignDevicesToRackResponse{AssignedCount: 2, RemovedCount: 1}), nil
+}
+
+func (h *recordingDeviceSetsHandler) SetRackSlotPosition(_ context.Context, request *connect.Request[devicesetv1.SetRackSlotPositionRequest]) (*connect.Response[devicesetv1.SetRackSlotPositionResponse], error) {
+	h.setRequests = append(h.setRequests, request.Msg)
+	return connect.NewResponse(&devicesetv1.SetRackSlotPositionResponse{
+		DeviceSetId: request.Msg.GetDeviceSetId(),
+		Slot: &devicesetv1.RackSlot{
+			DeviceIdentifier: request.Msg.GetDeviceIdentifier(),
+			Position:         request.Msg.GetPosition(),
+		},
+	}), nil
+}
+
+func (h *recordingDeviceSetsHandler) ClearRackSlotPosition(_ context.Context, request *connect.Request[devicesetv1.ClearRackSlotPositionRequest]) (*connect.Response[devicesetv1.ClearRackSlotPositionResponse], error) {
+	h.clearRequests = append(h.clearRequests, request.Msg)
+	return connect.NewResponse(&devicesetv1.ClearRackSlotPositionResponse{}), nil
+}
+
+func (h *recordingDeviceSetsHandler) GetRackSlots(context.Context, *connect.Request[devicesetv1.GetRackSlotsRequest]) (*connect.Response[devicesetv1.GetRackSlotsResponse], error) {
+	return connect.NewResponse(&devicesetv1.GetRackSlotsResponse{Slots: h.slots}), nil
+}
+
+func (h staticPoolsHandler) ListPools(context.Context, *connect.Request[poolsv1.ListPoolsRequest]) (*connect.Response[poolsv1.ListPoolsResponse], error) {
+	return connect.NewResponse(&poolsv1.ListPoolsResponse{Pools: h.pools}), nil
+}
+
+type recordingCommandHandler struct {
+	capabilityRequest  *minercommandv1.CheckCommandCapabilitiesRequest
+	rebootRequest      *minercommandv1.RebootRequest
+	stopRequest        *minercommandv1.StopMiningRequest
+	startRequest       *minercommandv1.StartMiningRequest
+	blinkRequest       *minercommandv1.BlinkLEDRequest
+	capabilityResponse *minercommandv1.CheckCommandCapabilitiesResponse
+}
+
+func (h *recordingCommandHandler) CheckCommandCapabilities(_ context.Context, request *connect.Request[minercommandv1.CheckCommandCapabilitiesRequest]) (*connect.Response[minercommandv1.CheckCommandCapabilitiesResponse], error) {
+	h.capabilityRequest = request.Msg
+	if h.capabilityResponse != nil {
+		return connect.NewResponse(h.capabilityResponse), nil
+	}
+	return connect.NewResponse(&minercommandv1.CheckCommandCapabilitiesResponse{}), nil
+}
+
+func (h *recordingCommandHandler) Reboot(_ context.Context, request *connect.Request[minercommandv1.RebootRequest]) (*connect.Response[minercommandv1.RebootResponse], error) {
+	h.rebootRequest = request.Msg
+	return connect.NewResponse(&minercommandv1.RebootResponse{BatchIdentifier: "batch-reboot"}), nil
+}
+
+func (h *recordingCommandHandler) StopMining(_ context.Context, request *connect.Request[minercommandv1.StopMiningRequest]) (*connect.Response[minercommandv1.StopMiningResponse], error) {
+	h.stopRequest = request.Msg
+	return connect.NewResponse(&minercommandv1.StopMiningResponse{BatchIdentifier: "batch-stop"}), nil
+}
+
+func (h *recordingCommandHandler) StartMining(_ context.Context, request *connect.Request[minercommandv1.StartMiningRequest]) (*connect.Response[minercommandv1.StartMiningResponse], error) {
+	h.startRequest = request.Msg
+	return connect.NewResponse(&minercommandv1.StartMiningResponse{BatchIdentifier: "batch-start"}), nil
+}
+
+func (h *recordingCommandHandler) BlinkLED(_ context.Context, request *connect.Request[minercommandv1.BlinkLEDRequest]) (*connect.Response[minercommandv1.BlinkLEDResponse], error) {
+	h.blinkRequest = request.Msg
+	return connect.NewResponse(&minercommandv1.BlinkLEDResponse{BatchIdentifier: "batch-blink"}), nil
+}
+
+type recordingScheduleHandler struct {
+	createRequest *schedulev1.CreateScheduleRequest
+}
+
+func (h *recordingScheduleHandler) CreateSchedule(_ context.Context, request *connect.Request[schedulev1.CreateScheduleRequest]) (*connect.Response[schedulev1.CreateScheduleResponse], error) {
+	h.createRequest = request.Msg
+	return connect.NewResponse(&schedulev1.CreateScheduleResponse{
+		Schedule: &schedulev1.Schedule{Id: 77, Name: request.Msg.GetName()},
+	}), nil
+}
+
+func TestListPoolsOnlyDisclosesNamesToModelProvider(t *testing.T) {
+	tools := NewFleetTools(nil, nil, staticPoolsHandler{pools: []*poolsv1.Pool{{
+		PoolId:   42,
+		PoolName: "Primary pool",
+		Url:      "stratum+tcp://pool.example.com:3333",
+		Username: "bc1q-wallet.worker-01",
+	}}}, nil, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "list_pools", json.RawMessage(`{}`))
+
+	require.NoError(t, err)
+	assert.JSONEq(t, `{"pools":[{"name":"Primary pool"}]}`, output.Content)
+	assert.NotContains(t, output.Content, "pool.example.com")
+	assert.NotContains(t, output.Content, "bc1q-wallet")
+	assert.NotContains(t, output.Content, "42")
+}
+
+func TestWriteToolDefinitionsRequireConfirmation(t *testing.T) {
+	tools := NewFleetTools(nil, &recordingSitesHandler{}, nil, &recordingDeviceSetsHandler{}, nil, nil)
+	requiresConfirmation := make(map[string]bool)
+	for _, definition := range tools.Definitions() {
+		requiresConfirmation[definition.Name] = definition.RequiresConfirmation
+	}
+
+	assert.True(t, requiresConfirmation["create_site"])
+	assert.True(t, requiresConfirmation["create_rack"])
+	assert.True(t, requiresConfirmation["move_miners_to_rack"])
+	assert.True(t, requiresConfirmation["set_rack_slots"])
+	assert.True(t, requiresConfirmation["clear_rack_slots"])
+	assert.True(t, requiresConfirmation["execute_miner_action"])
+	assert.True(t, requiresConfirmation["create_downtime_window"])
+	assert.False(t, requiresConfirmation["resolve_miners"])
+	assert.False(t, requiresConfirmation["list_sites"])
+	assert.False(t, requiresConfirmation["get_rack_slots"])
+	assert.False(t, requiresConfirmation["get_rack_health"])
+	assert.False(t, requiresConfirmation["get_site_health_summary"])
+	assert.False(t, requiresConfirmation["list_actionable_miner_issues"])
+	assert.False(t, requiresConfirmation["preview_miner_action"])
+	assert.False(t, requiresConfirmation["preview_downtime_window"])
+}
+
+func TestResolveMinersReturnsExplicitIdentifiersAndPlacement(t *testing.T) {
+	fleet := &staticFleetHandler{
+		totalMiners: 2,
+		miners: []*fleetv1.MinerStateSnapshot{
+			{
+				DeviceIdentifier: "miner-a",
+				Name:             "Alpha 01",
+				MacAddress:       "aa:bb:cc:dd:ee:ff",
+				SerialNumber:     "SN-ALPHA",
+				IpAddress:        "10.0.0.10",
+				DeviceStatus:     fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE,
+				Model:            "S21",
+				Placement: &commonv1.PlacementRefs{
+					Site:     &commonv1.ResourceRef{Id: 9, Label: "North"},
+					Building: &commonv1.ResourceRef{Id: 3, Label: "Building A"},
+					Rack:     &commonv1.ResourceRef{Id: 21, Label: "A1"},
+					Zone:     "Hot aisle",
+				},
+			},
+			{
+				DeviceIdentifier: "miner-b",
+				Name:             "Beta 01",
+				DeviceStatus:     fleetv1.DeviceStatus_DEVICE_STATUS_ONLINE,
+			},
+		},
+	}
+	tools := NewFleetTools(fleet, nil, nil, nil, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "resolve_miners", json.RawMessage(`{
+		"query":"alpha",
+		"device_statuses":["offline"],
+		"site_ids":[9],
+		"include_no_rack":true,
+		"limit":1
+	}`))
+
+	require.NoError(t, err)
+	require.Len(t, fleet.listRequests, 1)
+	request := fleet.listRequests[0]
+	assert.Equal(t, int32(1000), request.GetPageSize(), "query resolution scans a full page")
+	assert.Equal(t, []fleetv1.DeviceStatus{fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE}, request.GetFilter().GetDeviceStatus())
+	assert.Equal(t, []int64{9}, request.GetFilter().GetSiteIds())
+	assert.True(t, request.GetFilter().GetIncludeNoRack())
+	assert.JSONEq(t, `{
+		"device_identifiers":["miner-a"],
+		"miners":[{
+			"device_identifier":"miner-a",
+			"name":"Alpha 01",
+			"status":"offline",
+			"model":"S21",
+			"ip_address":"10.0.0.10",
+			"site":{"id":9,"label":"North"},
+			"building":{"id":3,"label":"Building A"},
+			"rack":{"id":21,"label":"A1"},
+			"zone":"Hot aisle"
+		}],
+		"returned":1,
+		"matched_scanned":1,
+		"total_available":2,
+		"truncated":false,
+		"query":"alpha"
+	}`, output.Content)
+	assert.NotContains(t, output.Content, "aa:bb:cc")
+	assert.NotContains(t, output.Content, "SN-ALPHA")
+	assert.Equal(t, "Resolved 1 miner(s)", output.Summary)
+}
+
+func TestResolveMinersReportsTruncatedMatches(t *testing.T) {
+	fleet := &staticFleetHandler{
+		totalMiners: 2,
+		miners: []*fleetv1.MinerStateSnapshot{
+			{DeviceIdentifier: "miner-a", DeviceStatus: fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE},
+			{DeviceIdentifier: "miner-b", DeviceStatus: fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE},
+		},
+		cursor: "next-page",
+	}
+	tools := NewFleetTools(fleet, nil, nil, nil, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "resolve_miners", json.RawMessage(`{"device_statuses":["broken"],"limit":1}`))
+
+	require.NoError(t, err)
+	assert.Contains(t, output.Summary, "more matches may exist")
+	assert.JSONEq(t, `{
+		"device_identifiers":["miner-a"],
+		"miners":[{"device_identifier":"miner-a","status":"offline"}],
+		"returned":1,
+		"matched_scanned":2,
+		"total_available":2,
+		"truncated":true
+	}`, output.Content)
+	require.Len(t, fleet.listRequests, 1)
+	assert.Equal(t, []fleetv1.DeviceStatus{fleetv1.DeviceStatus_DEVICE_STATUS_ERROR}, fleet.listRequests[0].GetFilter().GetDeviceStatus())
+}
+
+func TestResolveMinersRejectsInvalidStatus(t *testing.T) {
+	tools := NewFleetTools(&staticFleetHandler{}, nil, nil, nil, nil, nil)
+
+	_, err := tools.Execute(t.Context(), "resolve_miners", json.RawMessage(`{"device_statuses":["reticulating"]}`))
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), `unsupported device_status "reticulating"`)
+}
+
+func TestGetSiteHealthSummaryReturnsCountsAndInventory(t *testing.T) {
+	fleet := &staticFleetHandler{
+		counts: &fleetv1.GetMinerStateCountsResponse{
+			TotalMiners: 14,
+			StateCounts: &telemetryv1.MinerStateCounts{
+				HashingCount:  8,
+				BrokenCount:   1,
+				OfflineCount:  3,
+				SleepingCount: 2,
+			},
+		},
+	}
+	sites := &recordingSitesHandler{
+		sites: []*sitesv1.SiteWithCounts{
+			{Site: &sitesv1.Site{Id: 9, Name: "North"}, DeviceCount: 14, BuildingCount: 2, RackCount: 3, InfrastructureDeviceCount: 1},
+			{Site: &sitesv1.Site{Id: 10, Name: "South"}, DeviceCount: 4},
+		},
+	}
+	tools := NewFleetTools(fleet, sites, nil, nil, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "get_site_health_summary", json.RawMessage(`{"site_id":9}`))
+
+	require.NoError(t, err)
+	require.Len(t, fleet.countRequests, 1)
+	assert.Equal(t, []int64{9}, fleet.countRequests[0].GetSiteIds())
+	assert.JSONEq(t, `{
+		"scope":{"site_id":9,"include_unassigned":false},
+		"total_miners":14,
+		"state_counts":{"hashing":8,"broken":1,"offline":3,"sleeping":2},
+		"sites":[{
+			"id":9,
+			"name":"North",
+			"device_count":14,
+			"building_count":2,
+			"rack_count":3,
+			"infrastructure_device_count":1
+		}]
+	}`, output.Content)
+}
+
+func TestListActionableMinerIssuesReturnsSuggestedActions(t *testing.T) {
+	fleet := &staticFleetHandler{
+		totalMiners: 2,
+		miners: []*fleetv1.MinerStateSnapshot{
+			{DeviceIdentifier: "miner-a", DeviceStatus: fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE, IpAddress: "10.0.0.10"},
+			{DeviceIdentifier: "miner-b", DeviceStatus: fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED},
+		},
+	}
+	tools := NewFleetTools(fleet, nil, nil, nil, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "list_actionable_miner_issues", json.RawMessage(`{"limit":2}`))
+
+	require.NoError(t, err)
+	require.Len(t, fleet.listRequests, 1)
+	assert.Equal(t, []fleetv1.DeviceStatus{
+		fleetv1.DeviceStatus_DEVICE_STATUS_OFFLINE,
+		fleetv1.DeviceStatus_DEVICE_STATUS_ERROR,
+		fleetv1.DeviceStatus_DEVICE_STATUS_NEEDS_MINING_POOL,
+		fleetv1.DeviceStatus_DEVICE_STATUS_REBOOT_REQUIRED,
+	}, fleet.listRequests[0].GetFilter().GetDeviceStatus())
+	assert.JSONEq(t, `{
+		"issues":[
+			{
+				"device_identifier":"miner-a",
+				"status":"offline",
+				"severity":"high",
+				"issue":"Miner is offline",
+				"suggested_action":"Check power and network reachability; if reachable, preview a reboot before executing it.",
+				"ip_address":"10.0.0.10"
+			},
+			{
+				"device_identifier":"miner-b",
+				"status":"reboot_required",
+				"severity":"medium",
+				"issue":"Miner reports reboot required",
+				"suggested_action":"Use preview_miner_action with reboot, then execute_miner_action after operator confirmation."
+			}
+		],
+		"returned":2,
+		"matched_scanned":2,
+		"total_available":2,
+		"truncated":false
+	}`, output.Content)
+}
+
+func TestCreateSiteConfirmationMatchesExecutedRequest(t *testing.T) {
+	sites := &recordingSitesHandler{}
+	tools := NewFleetTools(nil, sites, nil, nil, nil, nil)
+	arguments := json.RawMessage(`{"name":"North","location_city":"Austin","location_state":"TX","country":"US","power_capacity_mw":12.5}`)
+
+	confirmation, err := tools.Confirmation("create_site", arguments)
+	require.NoError(t, err)
+	require.NotNil(t, confirmation)
+	assert.Equal(t, "Create this site?", confirmation.Title)
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Name", Value: "North"})
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Power capacity", Value: "12.5 MW"})
+
+	output, err := tools.Execute(t.Context(), "create_site", arguments)
+	require.NoError(t, err)
+	require.NotNil(t, sites.createRequest)
+	assert.Equal(t, "North", sites.createRequest.GetName())
+	assert.Equal(t, 12.5, sites.createRequest.GetPowerCapacityMw())
+	assert.JSONEq(t, `{"created":true,"site_id":12,"name":"North","warnings":[]}`, output.Content)
+}
+
+func TestCreateRackAndMoveMinersUseValidatedExplicitInputs(t *testing.T) {
+	deviceSets := &recordingDeviceSetsHandler{}
+	tools := NewFleetTools(nil, nil, nil, deviceSets, nil, nil)
+
+	confirmation, err := tools.Confirmation("create_rack", json.RawMessage(`{"label":"A1","rows":4,"columns":6,"site_id":9}`))
+	require.NoError(t, err)
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Layout", Value: "4 rows × 6 columns"})
+	_, err = tools.Execute(t.Context(), "create_rack", json.RawMessage(`{"label":"A1","rows":4,"columns":6,"site_id":9}`))
+	require.NoError(t, err)
+	require.NotNil(t, deviceSets.createRequest)
+	assert.Equal(t, int64(9), deviceSets.createRequest.GetRackInfo().GetSiteId())
+
+	moveArguments := json.RawMessage(`{"target_rack_id":21,"device_identifiers":["miner-a","miner-b"]}`)
+	_, err = tools.Execute(t.Context(), "move_miners_to_rack", moveArguments)
+	require.NoError(t, err)
+	require.NotNil(t, deviceSets.moveRequest)
+	assert.Equal(t, int64(21), deviceSets.moveRequest.GetTargetRackId())
+	assert.Equal(t, []string{"miner-a", "miner-b"}, deviceSets.moveRequest.GetDeviceSelector().GetDeviceList().GetDeviceIdentifiers())
+}
+
+func TestGetRackSlotsReturnsOccupiedSlots(t *testing.T) {
+	deviceSets := &recordingDeviceSetsHandler{
+		slots: []*devicesetv1.RackSlot{
+			{DeviceIdentifier: "miner-a", Position: &devicesetv1.RackSlotPosition{Row: 0, Column: 1}},
+			{DeviceIdentifier: "miner-b", Position: &devicesetv1.RackSlotPosition{Row: 1, Column: 0}},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, deviceSets, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "get_rack_slots", json.RawMessage(`{"rack_id":21}`))
+
+	require.NoError(t, err)
+	assert.JSONEq(t, `{
+		"rack_id":21,
+		"occupied_count":2,
+		"occupied_slots":[
+			{"device_identifier":"miner-a","row":0,"column":1},
+			{"device_identifier":"miner-b","row":1,"column":0}
+		]
+	}`, output.Content)
+	assert.Equal(t, "Read 2 occupied slot(s) for rack 21", output.Summary)
+}
+
+func TestGetRackHealthReturnsLayoutStatusAndSlots(t *testing.T) {
+	fleet := &staticFleetHandler{
+		totalMiners: 2,
+		miners: []*fleetv1.MinerStateSnapshot{
+			{
+				DeviceIdentifier: "miner-a",
+				Name:             "Alpha",
+				DeviceStatus:     fleetv1.DeviceStatus_DEVICE_STATUS_ONLINE,
+				Placement: &commonv1.PlacementRefs{
+					Rack: &commonv1.ResourceRef{Id: 21, Label: "A1"},
+				},
+			},
+			{
+				DeviceIdentifier: "miner-b",
+				DeviceStatus:     fleetv1.DeviceStatus_DEVICE_STATUS_ERROR,
+			},
+		},
+	}
+	deviceSets := &recordingDeviceSetsHandler{
+		rack: &devicesetv1.DeviceSet{
+			Id:          21,
+			Type:        devicesetv1.DeviceSetType_DEVICE_SET_TYPE_RACK,
+			Label:       "A1",
+			DeviceCount: 2,
+			TypeDetails: &devicesetv1.DeviceSet_RackInfo{RackInfo: &devicesetv1.RackInfo{
+				Rows:       2,
+				Columns:    2,
+				OrderIndex: devicesetv1.RackOrderIndex_RACK_ORDER_INDEX_BOTTOM_LEFT,
+				Zone:       "Z1",
+			}},
+		},
+		members: []*devicesetv1.DeviceSetMember{
+			{DeviceIdentifier: "miner-a"},
+			{DeviceIdentifier: "miner-b"},
+		},
+		slots: []*devicesetv1.RackSlot{
+			{DeviceIdentifier: "miner-a", Position: &devicesetv1.RackSlotPosition{Row: 1, Column: 0}},
+		},
+	}
+	tools := NewFleetTools(fleet, nil, nil, deviceSets, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "get_rack_health", json.RawMessage(`{"rack_id":21}`))
+
+	require.NoError(t, err)
+	require.Len(t, fleet.listRequests, 1)
+	assert.Equal(t, []int64{21}, fleet.listRequests[0].GetFilter().GetRackIds())
+	assert.JSONEq(t, `{
+		"rack":{"id":21,"label":"A1","rows":2,"columns":2,"numbering_origin":"bottom_left","zone":"Z1","device_count":2},
+		"member_count":2,
+		"occupied_slot_count":1,
+		"status_counts":{"hashing":1,"broken":1,"offline":0,"sleeping":0,"maintenance":0,"needs_mining_pool":0,"updating":0,"reboot_required":0,"status_unspecified":0},
+		"miners":[
+			{"device_identifier":"miner-a","name":"Alpha","status":"online","rack":{"id":21,"label":"A1"},"slot":{"row":1,"column":0}},
+			{"device_identifier":"miner-b","status":"error"}
+		],
+		"returned":2,
+		"total_available":2,
+		"truncated":false
+	}`, output.Content)
+}
+
+func TestSetRackSlotsClearsThenAssignsRequestedSlots(t *testing.T) {
+	deviceSets := &recordingDeviceSetsHandler{
+		members: []*devicesetv1.DeviceSetMember{
+			{DeviceIdentifier: "miner-a"},
+			{DeviceIdentifier: "miner-b"},
+			{DeviceIdentifier: "miner-c"},
+		},
+		slots: []*devicesetv1.RackSlot{
+			{DeviceIdentifier: "miner-a", Position: &devicesetv1.RackSlotPosition{Row: 0, Column: 0}},
+			{DeviceIdentifier: "miner-c", Position: &devicesetv1.RackSlotPosition{Row: 2, Column: 0}},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, deviceSets, nil, nil)
+	arguments := json.RawMessage(`{"rack_id":21,"slot_assignments":[
+		{"device_identifier":"miner-a","row":0,"column":1},
+		{"device_identifier":"miner-b","row":0,"column":0}
+	]}`)
+
+	confirmation, err := tools.Confirmation("set_rack_slots", arguments)
+	require.NoError(t, err)
+	assert.Equal(t, "Assign 2 rack slot(s)?", confirmation.Title)
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Rack ID", Value: "21"})
+
+	output, err := tools.Execute(t.Context(), "set_rack_slots", arguments)
+
+	require.NoError(t, err)
+	require.Len(t, deviceSets.clearRequests, 2)
+	assert.ElementsMatch(t, []string{"miner-a", "miner-b"}, []string{
+		deviceSets.clearRequests[0].GetDeviceIdentifier(),
+		deviceSets.clearRequests[1].GetDeviceIdentifier(),
+	})
+	require.Len(t, deviceSets.setRequests, 2)
+	assert.Equal(t, int64(21), deviceSets.setRequests[0].GetDeviceSetId())
+	assert.Equal(t, "miner-a", deviceSets.setRequests[0].GetDeviceIdentifier())
+	assert.Equal(t, int32(0), deviceSets.setRequests[0].GetPosition().GetRow())
+	assert.Equal(t, int32(1), deviceSets.setRequests[0].GetPosition().GetColumn())
+	assert.Equal(t, "miner-b", deviceSets.setRequests[1].GetDeviceIdentifier())
+	assert.Equal(t, int32(0), deviceSets.setRequests[1].GetPosition().GetRow())
+	assert.Equal(t, int32(0), deviceSets.setRequests[1].GetPosition().GetColumn())
+	assert.JSONEq(t, `{
+		"applied":true,
+		"rack_id":21,
+		"rack_label":"A1",
+		"assigned_count":2,
+		"slot_assignments":[
+			{"device_identifier":"miner-a","row":0,"column":1},
+			{"device_identifier":"miner-b","row":0,"column":0}
+		]
+	}`, output.Content)
+}
+
+func TestSetRackSlotsRejectsUnrequestedOccupiedSlot(t *testing.T) {
+	deviceSets := &recordingDeviceSetsHandler{
+		members: []*devicesetv1.DeviceSetMember{
+			{DeviceIdentifier: "miner-a"},
+			{DeviceIdentifier: "miner-b"},
+		},
+		slots: []*devicesetv1.RackSlot{
+			{DeviceIdentifier: "miner-b", Position: &devicesetv1.RackSlotPosition{Row: 0, Column: 1}},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, deviceSets, nil, nil)
+
+	_, err := tools.Execute(t.Context(), "set_rack_slots", json.RawMessage(`{"rack_id":21,"slot_assignments":[
+		{"device_identifier":"miner-a","row":0,"column":1}
+	]}`))
+
+	require.Error(t, err)
+	assert.Contains(t, err.Error(), `slot (0,1) is already occupied by miner "miner-b"`)
+	assert.Empty(t, deviceSets.clearRequests)
+	assert.Empty(t, deviceSets.setRequests)
+}
+
+func TestClearRackSlotsClearsOnlyRequestedMiners(t *testing.T) {
+	deviceSets := &recordingDeviceSetsHandler{
+		members: []*devicesetv1.DeviceSetMember{
+			{DeviceIdentifier: "miner-a"},
+			{DeviceIdentifier: "miner-b"},
+		},
+		slots: []*devicesetv1.RackSlot{
+			{DeviceIdentifier: "miner-a", Position: &devicesetv1.RackSlotPosition{Row: 0, Column: 1}},
+			{DeviceIdentifier: "miner-b", Position: &devicesetv1.RackSlotPosition{Row: 1, Column: 1}},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, deviceSets, nil, nil)
+
+	output, err := tools.Execute(t.Context(), "clear_rack_slots", json.RawMessage(`{"rack_id":21,"device_identifiers":["miner-a"]}`))
+
+	require.NoError(t, err)
+	require.Len(t, deviceSets.clearRequests, 1)
+	assert.Equal(t, int64(21), deviceSets.clearRequests[0].GetDeviceSetId())
+	assert.Equal(t, "miner-a", deviceSets.clearRequests[0].GetDeviceIdentifier())
+	assert.Empty(t, deviceSets.setRequests)
+	assert.JSONEq(t, `{
+		"cleared":true,
+		"rack_id":21,
+		"rack_label":"A1",
+		"requested_count":1,
+		"cleared_count":1,
+		"device_identifiers":["miner-a"]
+	}`, output.Content)
+}
+
+func TestPreviewAndExecuteMinerActionUseExistingCommandHandler(t *testing.T) {
+	commands := &recordingCommandHandler{
+		capabilityResponse: &minercommandv1.CheckCommandCapabilitiesResponse{
+			TotalCount:                 2,
+			SupportedCount:             1,
+			UnsupportedCount:           1,
+			AllSupported:               false,
+			SupportedDeviceIdentifiers: []string{"miner-a"},
+			UnsupportedGroups: []*minercommandv1.UnsupportedMinerGroup{
+				{Model: "S19", FirmwareVersion: "1.0.0", Count: 1},
+			},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, nil, commands, nil)
+	previewArguments := json.RawMessage(`{"action":"reboot","device_identifiers":["miner-a","miner-b"]}`)
+
+	preview, err := tools.Execute(t.Context(), "preview_miner_action", previewArguments)
+
+	require.NoError(t, err)
+	require.NotNil(t, commands.capabilityRequest)
+	assert.Equal(t, minercommandv1.CommandType_COMMAND_TYPE_REBOOT, commands.capabilityRequest.GetCommandType())
+	assert.Equal(t, []string{"miner-a", "miner-b"}, commands.capabilityRequest.GetDeviceSelector().GetIncludeDevices().GetDeviceIdentifiers())
+	assert.JSONEq(t, `{
+		"action":"reboot",
+		"selector":{"type":"explicit","description":"2 explicit miner(s)","device_identifiers":["miner-a","miner-b"]},
+		"requested_device_identifiers":["miner-a","miner-b"],
+		"total_count":2,
+		"supported_count":1,
+		"unsupported_count":1,
+		"all_supported":false,
+		"none_supported":false,
+		"supported_device_identifiers":["miner-a"],
+		"unsupported_groups":[{"firmware_version":"1.0.0","model":"S19","count":1}]
+	}`, preview.Content)
+
+	confirmation, err := tools.Confirmation("execute_miner_action", previewArguments)
+	require.NoError(t, err)
+	assert.Equal(t, "Execute reboot on 2 miner(s)?", confirmation.Title)
+
+	output, err := tools.Execute(t.Context(), "execute_miner_action", previewArguments)
+
+	require.NoError(t, err)
+	require.NotNil(t, commands.rebootRequest)
+	assert.Equal(t, []string{"miner-a", "miner-b"}, commands.rebootRequest.GetDeviceSelector().GetIncludeDevices().GetDeviceIdentifiers())
+	assert.JSONEq(t, `{
+		"executed":true,
+		"action":"reboot",
+		"selector":{"type":"explicit","description":"2 explicit miner(s)","device_identifiers":["miner-a","miner-b"]},
+		"selected_count":2,
+		"device_identifiers":["miner-a","miner-b"],
+		"command_batch_identifier":"batch-reboot",
+		"command_status_lookup_hint":"Use the existing command activity UI to inspect per-miner command results."
+	}`, output.Content)
+}
+
+func TestMinerActionSupportsWholeFleetSelectorWithoutEnumeratingIDs(t *testing.T) {
+	commands := &recordingCommandHandler{
+		capabilityResponse: &minercommandv1.CheckCommandCapabilitiesResponse{
+			TotalCount:                 14,
+			SupportedCount:             14,
+			AllSupported:               true,
+			SupportedDeviceIdentifiers: []string{"miner-a", "miner-b"},
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, nil, commands, nil)
+	arguments := json.RawMessage(`{"action":"stop_mining","selector":{"type":"all_devices"}}`)
+
+	preview, err := tools.Execute(t.Context(), "preview_miner_action", arguments)
+
+	require.NoError(t, err)
+	require.NotNil(t, commands.capabilityRequest.GetDeviceSelector().GetAllDevices())
+	assert.JSONEq(t, `{
+		"action":"stop_mining",
+		"selector":{"type":"all_devices","description":"whole fleet command-eligible miners"},
+		"total_count":14,
+		"supported_count":14,
+		"unsupported_count":0,
+		"all_supported":true,
+		"none_supported":false,
+		"supported_device_identifiers":["miner-a","miner-b"],
+		"unsupported_groups":[]
+	}`, preview.Content)
+
+	confirmation, err := tools.Confirmation("execute_miner_action", arguments)
+	require.NoError(t, err)
+	assert.Equal(t, "Execute stop mining on the whole fleet?", confirmation.Title)
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Selection", Value: "Whole fleet command-eligible miners"})
+
+	output, err := tools.Execute(t.Context(), "execute_miner_action", arguments)
+
+	require.NoError(t, err)
+	require.NotNil(t, commands.stopRequest.GetDeviceSelector().GetAllDevices())
+	assert.JSONEq(t, `{
+		"executed":true,
+		"action":"stop_mining",
+		"selector":{"type":"all_devices","description":"whole fleet command-eligible miners"},
+		"command_batch_identifier":"batch-stop",
+		"command_status_lookup_hint":"Use the existing command activity UI to inspect per-miner command results."
+	}`, output.Content)
+}
+
+func TestMinerActionSupportsBackendFilterSelector(t *testing.T) {
+	commands := &recordingCommandHandler{
+		capabilityResponse: &minercommandv1.CheckCommandCapabilitiesResponse{
+			TotalCount:     3,
+			SupportedCount: 3,
+			AllSupported:   true,
+		},
+	}
+	tools := NewFleetTools(nil, nil, nil, nil, commands, nil)
+	arguments := json.RawMessage(`{
+		"action":"start_mining",
+		"selector":{
+			"type":"filter",
+			"filter":{
+				"device_statuses":["sleeping"],
+				"site_ids":[9],
+				"rack_ids":[21],
+				"models":["S21"]
+			}
+		}
+	}`)
+
+	_, err := tools.Execute(t.Context(), "preview_miner_action", arguments)
+
+	require.NoError(t, err)
+	filter := commands.capabilityRequest.GetDeviceSelector().GetAllMatchingFilter()
+	require.NotNil(t, filter)
+	assert.Equal(t, []fleetv1.DeviceStatus{fleetv1.DeviceStatus_DEVICE_STATUS_INACTIVE}, filter.GetDeviceStatus())
+	assert.Equal(t, []int64{9}, filter.GetSiteIds())
+	assert.Equal(t, []int64{21}, filter.GetRackIds())
+	assert.Equal(t, []string{"S21"}, filter.GetModels())
+
+	output, err := tools.Execute(t.Context(), "execute_miner_action", arguments)
+
+	require.NoError(t, err)
+	filter = commands.startRequest.GetDeviceSelector().GetAllMatchingFilter()
+	require.NotNil(t, filter)
+	assert.Equal(t, []int64{9}, filter.GetSiteIds())
+	assert.JSONEq(t, `{
+		"executed":true,
+		"action":"start_mining",
+		"selector":{
+			"type":"filter",
+			"description":"miners matching the supplied filter",
+			"filter":{
+				"device_statuses":["sleeping"],
+				"site_ids":[9],
+				"include_unassigned":false,
+				"building_ids":null,
+				"include_no_building":false,
+				"rack_ids":[21],
+				"include_no_rack":false,
+				"models":["S21"],
+				"ip_cidrs":null
+			}
+		},
+		"command_batch_identifier":"batch-start",
+		"command_status_lookup_hint":"Use the existing command activity UI to inspect per-miner command results."
+	}`, output.Content)
+}
+
+func TestCreateDowntimeWindowUsesExistingScheduleHandler(t *testing.T) {
+	schedules := &recordingScheduleHandler{}
+	tools := NewFleetTools(nil, nil, nil, nil, nil, schedules)
+	arguments := json.RawMessage(`{
+		"name":"Rack A maintenance",
+		"action":"sleep",
+		"start_date":"2026-08-01",
+		"start_time":"02:00",
+		"end_time":"04:00",
+		"timezone":"America/Chicago",
+		"targets":[{"type":"rack","target_id":"21"}]
+	}`)
+
+	preview, err := tools.Execute(t.Context(), "preview_downtime_window", arguments)
+	require.NoError(t, err)
+	assert.JSONEq(t, `{
+		"will_create":false,
+		"schedule_id":0,
+		"name":"Rack A maintenance",
+		"action":"sleep",
+		"schedule_type":"one_time",
+		"start_date":"2026-08-01",
+		"start_time":"02:00",
+		"end_date":"",
+		"end_time":"04:00",
+		"timezone":"America/Chicago",
+		"targets":[{"type":"rack","target_id":"21"}],
+		"target_count":1,
+		"operator_note":"Sleep schedules stop mining at start_time and do not automatically resume at end_time."
+	}`, preview.Content)
+
+	confirmation, err := tools.Confirmation("create_downtime_window", arguments)
+	require.NoError(t, err)
+	assert.Equal(t, "Create schedule \"Rack A maintenance\"?", confirmation.Title)
+	assert.Contains(t, confirmation.Details, chatdomain.ToolConfirmationDetail{Label: "Resume", Value: "Not automatic for sleep schedules"})
+
+	output, err := tools.Execute(t.Context(), "create_downtime_window", arguments)
+
+	require.NoError(t, err)
+	require.NotNil(t, schedules.createRequest)
+	assert.Equal(t, "Rack A maintenance", schedules.createRequest.GetName())
+	assert.Equal(t, schedulev1.ScheduleAction_SCHEDULE_ACTION_SLEEP, schedules.createRequest.GetAction())
+	assert.Equal(t, schedulev1.ScheduleType_SCHEDULE_TYPE_ONE_TIME, schedules.createRequest.GetScheduleType())
+	require.Len(t, schedules.createRequest.GetTargets(), 1)
+	assert.Equal(t, schedulev1.ScheduleTargetType_SCHEDULE_TARGET_TYPE_RACK, schedules.createRequest.GetTargets()[0].GetTargetType())
+	assert.Equal(t, "21", schedules.createRequest.GetTargets()[0].GetTargetId())
+	assert.JSONEq(t, `{
+		"created":true,
+		"schedule_id":77,
+		"name":"Rack A maintenance",
+		"action":"sleep",
+		"schedule_type":"one_time",
+		"start_date":"2026-08-01",
+		"start_time":"02:00",
+		"end_date":"",
+		"end_time":"04:00",
+		"timezone":"America/Chicago",
+		"targets":[{"type":"rack","target_id":"21"}],
+		"target_count":1,
+		"operator_note":"Sleep schedules stop mining at start_time and do not automatically resume at end_time."
+	}`, output.Content)
+}
+
+func TestWriteConfirmationRejectsArgumentsThatWouldNotBeExecuted(t *testing.T) {
+	tools := NewFleetTools(nil, &recordingSitesHandler{}, nil, nil, nil, nil)
+
+	confirmation, err := tools.Confirmation("create_site", json.RawMessage(`{"name":"North","unexpected":true}`))
+
+	require.Error(t, err)
+	assert.Nil(t, confirmation)
+}
diff --git a/server/internal/handlers/interceptors/config.go b/server/internal/handlers/interceptors/config.go
index 947385c75..e36fe898b 100644
--- a/server/internal/handlers/interceptors/config.go
+++ b/server/internal/handlers/interceptors/config.go
@@ -5,6 +5,7 @@ import (
 	"github.com/block/proto-fleet/server/generated/grpc/apikey/v1/apikeyv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/auth/v1/authv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/authz/v1/authzv1connect"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/curtailment/v1/curtailmentv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/fleetmanagement/v1/fleetmanagementv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/fleetnodeadmin/v1/fleetnodeadminv1connect"
@@ -48,6 +49,8 @@ var RedactedRequestProcedures = []string{
 	alertsv1connect.ChannelServiceTestChannelProcedure,
 	// Site control-subnet commissioning carries OT topology.
 	sitesv1connect.SiteServiceSetInfrastructureControlSubnetsProcedure,
+	chatv1connect.ChatServiceDiscoverModelsProcedure,
+	chatv1connect.ChatServiceUpdateLLMConfigProcedure,
 }
 
 // RedactedResponseProcedures lists procedures whose responses contain secrets
@@ -154,6 +157,14 @@ var SessionOnlyProcedures = []string{
 	alertsv1connect.MaintenanceWindowServiceUpdateMaintenanceWindowProcedure,
 	alertsv1connect.MaintenanceWindowServiceDeleteMaintenanceWindowProcedure,
 	alertsv1connect.HistoryServiceListAlertsProcedure,
+	// Provider settings and interactive prompts can contain credentials or
+	// operationally sensitive fleet context. Keep the full AI surface bound to
+	// an authenticated browser session rather than reusable API keys.
+	chatv1connect.ChatServiceGetLLMConfigProcedure,
+	chatv1connect.ChatServiceDiscoverModelsProcedure,
+	chatv1connect.ChatServiceUpdateLLMConfigProcedure,
+	chatv1connect.ChatServiceSendMessageProcedure,
+	chatv1connect.ChatServiceResolveToolConfirmationProcedure,
 }
 
 var UnauthenticatedProcedures = []string{
@@ -202,6 +213,9 @@ var SensitiveBodyProcedures = map[string]bool{
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceUploadTelemetryProcedure:         true,
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceUploadEventsProcedure:            true,
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceUploadCommandArtifactProcedure:   true,
+	chatv1connect.ChatServiceUpdateLLMConfigProcedure:                                 true,
+	chatv1connect.ChatServiceDiscoverModelsProcedure:                                  true,
+	chatv1connect.ChatServiceSendMessageProcedure:                                     true,
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceDownloadCommandArtifactProcedure: true,
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceReportDiscoveredDevicesProcedure: true,
 	fleetnodegatewayv1connect.FleetNodeGatewayServiceReportPairedDevicesProcedure:     true,
diff --git a/server/internal/handlers/interceptors/config_test.go b/server/internal/handlers/interceptors/config_test.go
index 6f7f3cb45..a9c696f30 100644
--- a/server/internal/handlers/interceptors/config_test.go
+++ b/server/internal/handlers/interceptors/config_test.go
@@ -11,6 +11,7 @@ import (
 
 	"github.com/block/proto-fleet/server/generated/grpc/alerts/v1/alertsv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/auth/v1/authv1connect"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/curtailment/v1/curtailmentv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/fleetmanagement/v1/fleetmanagementv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/infrastructure/v1/infrastructurev1connect"
@@ -49,6 +50,33 @@ func TestMqttSettingsPasswordProceduresAreRedacted(t *testing.T) {
 	}
 }
 
+func TestChatProceduresProtectSecretsAndPromptBodies(t *testing.T) {
+	t.Parallel()
+
+	configProcedure := chatv1connect.ChatServiceUpdateLLMConfigProcedure
+	assert.Contains(t, RedactedRequestProcedures, configProcedure)
+	assert.True(t, SensitiveBodyProcedures[configProcedure])
+	discoveryProcedure := chatv1connect.ChatServiceDiscoverModelsProcedure
+	assert.Contains(t, RedactedRequestProcedures, discoveryProcedure)
+	assert.True(t, SensitiveBodyProcedures[discoveryProcedure])
+
+	chatProcedure := chatv1connect.ChatServiceSendMessageProcedure
+	assert.True(t, SensitiveBodyProcedures[chatProcedure])
+	confirmationProcedure := chatv1connect.ChatServiceResolveToolConfirmationProcedure
+
+	for _, procedure := range []string{
+		chatv1connect.ChatServiceGetLLMConfigProcedure,
+		discoveryProcedure,
+		configProcedure,
+		chatProcedure,
+		confirmationProcedure,
+	} {
+		assert.Contains(t, SessionOnlyProcedures, procedure,
+			"%s must reject API-key auth because the AI surface contains credentials or fleet context",
+			procedure)
+	}
+}
+
 // Infrastructure device bodies carry driver_config — the OT control
 // network map (endpoint IPs, unit IDs, register addresses) — and must
 // never land in debug logs.
diff --git a/server/internal/handlers/middleware/rpc_permissions.go b/server/internal/handlers/middleware/rpc_permissions.go
index 44fccf5bb..2b58b72fb 100644
--- a/server/internal/handlers/middleware/rpc_permissions.go
+++ b/server/internal/handlers/middleware/rpc_permissions.go
@@ -7,6 +7,7 @@ import (
 	"github.com/block/proto-fleet/server/generated/grpc/auth/v1/authv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/authz/v1/authzv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/buildings/v1/buildingsv1connect"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/collection/v1/collectionv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/curtailment/v1/curtailmentv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/device_set/v1/device_setv1connect"
@@ -49,6 +50,15 @@ import (
 // glance: shrinking ProceduresPendingMigration to zero is the exit
 // criterion for retiring the legacy RequireAdmin middleware.
 var ProcedurePermissions = map[string]string{
+	// AI provider settings share the existing integrations-management gate;
+	// conversations require fleet read and each invoked tool applies its own
+	// narrower handler permission before returning data.
+	chatv1connect.ChatServiceGetLLMConfigProcedure:            authz.PermAPIKeyManage,
+	chatv1connect.ChatServiceDiscoverModelsProcedure:          authz.PermAPIKeyManage,
+	chatv1connect.ChatServiceUpdateLLMConfigProcedure:         authz.PermAPIKeyManage,
+	chatv1connect.ChatServiceSendMessageProcedure:             authz.PermFleetRead,
+	chatv1connect.ChatServiceResolveToolConfirmationProcedure: authz.PermFleetRead,
+
 	// Activity log — read-only audit trail. Export is the CSV variant of
 	// the same query; filter options is the lookup endpoint that drives
 	// the UI's filter panel. All three sit on activity:read.
diff --git a/server/internal/handlers/middleware/rpc_permissions_test.go b/server/internal/handlers/middleware/rpc_permissions_test.go
index 5592bbe0e..80ff8a615 100644
--- a/server/internal/handlers/middleware/rpc_permissions_test.go
+++ b/server/internal/handlers/middleware/rpc_permissions_test.go
@@ -19,6 +19,7 @@ import (
 	"github.com/block/proto-fleet/server/generated/grpc/auth/v1/authv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/authz/v1/authzv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/buildings/v1/buildingsv1connect"
+	"github.com/block/proto-fleet/server/generated/grpc/chat/v1/chatv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/collection/v1/collectionv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/curtailment/v1/curtailmentv1connect"
 	"github.com/block/proto-fleet/server/generated/grpc/device_set/v1/device_setv1connect"
@@ -70,6 +71,7 @@ var registeredServices = []struct {
 	{authv1connect.AuthServiceName, reflect.TypeOf((*authv1connect.AuthServiceHandler)(nil)).Elem()},
 	{authzv1connect.AuthzServiceName, reflect.TypeOf((*authzv1connect.AuthzServiceHandler)(nil)).Elem()},
 	{buildingsv1connect.BuildingServiceName, reflect.TypeOf((*buildingsv1connect.BuildingServiceHandler)(nil)).Elem()},
+	{chatv1connect.ChatServiceName, reflect.TypeOf((*chatv1connect.ChatServiceHandler)(nil)).Elem()},
 	{collectionv1connect.DeviceCollectionServiceName, reflect.TypeOf((*collectionv1connect.DeviceCollectionServiceHandler)(nil)).Elem()},
 	{curtailmentv1connect.CurtailmentServiceName, reflect.TypeOf((*curtailmentv1connect.CurtailmentServiceHandler)(nil)).Elem()},
 	{device_setv1connect.DeviceSetServiceName, reflect.TypeOf((*device_setv1connect.DeviceSetServiceHandler)(nil)).Elem()},
diff --git a/server/migrations/000129_create_llm_config.down.sql b/server/migrations/000129_create_llm_config.down.sql
new file mode 100644
index 000000000..9eb254355
--- /dev/null
+++ b/server/migrations/000129_create_llm_config.down.sql
@@ -0,0 +1 @@
+DROP TABLE IF EXISTS llm_config;
diff --git a/server/migrations/000129_create_llm_config.up.sql b/server/migrations/000129_create_llm_config.up.sql
new file mode 100644
index 000000000..1703dd81b
--- /dev/null
+++ b/server/migrations/000129_create_llm_config.up.sql
@@ -0,0 +1,17 @@
+CREATE TABLE llm_config (
+    organization_id         BIGINT PRIMARY KEY REFERENCES organization(id) ON DELETE CASCADE,
+    harness                 TEXT NOT NULL CHECK (harness IN ('native', 'goose')),
+    provider                TEXT NOT NULL CHECK (provider IN ('openai', 'anthropic', 'ollama', 'custom')),
+    api_key_encrypted       TEXT NOT NULL DEFAULT '',
+    base_url                TEXT NOT NULL DEFAULT '',
+    model                   TEXT NOT NULL,
+    temperature             DOUBLE PRECISION NOT NULL DEFAULT 0.2 CHECK (temperature >= 0 AND temperature <= 1),
+    goose_base_url          TEXT NOT NULL DEFAULT '',
+    goose_secret_encrypted  TEXT NOT NULL DEFAULT '',
+    created_at              TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP,
+    updated_at              TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP
+);
+
+COMMENT ON TABLE llm_config IS 'Organization-scoped agent harness and encrypted BYOLLM provider configuration.';
+COMMENT ON COLUMN llm_config.api_key_encrypted IS 'AES-GCM ciphertext produced by the fleet service master key; never returned over RPC.';
+COMMENT ON COLUMN llm_config.goose_secret_encrypted IS 'Encrypted authentication secret for a future remote Goose ACP adapter.';
diff --git a/server/sqlc/queries/llm_config.sql b/server/sqlc/queries/llm_config.sql
new file mode 100644
index 000000000..ad9b516e2
--- /dev/null
+++ b/server/sqlc/queries/llm_config.sql
@@ -0,0 +1,60 @@
+-- name: GetLLMConfig :one
+SELECT
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted,
+    created_at,
+    updated_at
+FROM llm_config
+WHERE organization_id = $1;
+
+-- name: UpsertLLMConfig :one
+INSERT INTO llm_config (
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted
+) VALUES (
+    sqlc.arg('organization_id'),
+    sqlc.arg('harness'),
+    sqlc.arg('provider'),
+    sqlc.arg('api_key_encrypted'),
+    sqlc.arg('base_url'),
+    sqlc.arg('model'),
+    sqlc.arg('temperature'),
+    sqlc.arg('goose_base_url'),
+    sqlc.arg('goose_secret_encrypted')
+)
+ON CONFLICT (organization_id) DO UPDATE SET
+    harness = EXCLUDED.harness,
+    provider = EXCLUDED.provider,
+    api_key_encrypted = EXCLUDED.api_key_encrypted,
+    base_url = EXCLUDED.base_url,
+    model = EXCLUDED.model,
+    temperature = EXCLUDED.temperature,
+    goose_base_url = EXCLUDED.goose_base_url,
+    goose_secret_encrypted = EXCLUDED.goose_secret_encrypted,
+    updated_at = CURRENT_TIMESTAMP
+RETURNING
+    organization_id,
+    harness,
+    provider,
+    api_key_encrypted,
+    base_url,
+    model,
+    temperature,
+    goose_base_url,
+    goose_secret_encrypted,
+    created_at,
+    updated_at;