diff --git a/src-tauri/src/agents/openai_compat.rs b/src-tauri/src/agents/openai_compat.rs
index 1e926f69..cd1997d0 100644
--- a/src-tauri/src/agents/openai_compat.rs
+++ b/src-tauri/src/agents/openai_compat.rs
@@ -87,6 +87,41 @@ pub fn lmstudio_base_url() -> String {
         .unwrap_or_else(|_| "http://localhost:1234".into())
 }
 
+/// vLLM base URL (default: localhost:8000)
+pub fn vllm_base_url() -> String {
+    std::env::var("VLLM_ENDPOINT")
+        .unwrap_or_else(|_| "http://localhost:8000".into())
+}
+
+/// Discover vLLM models via OpenAI-compatible `/v1/models` endpoint.
+pub fn discover_vllm() -> Option<Vec<String>> {
+    let endpoint = vllm_base_url();
+    let url = format!("{}/v1/models", endpoint.trim_end_matches('/'));
+
+    let client = reqwest::blocking::Client::builder()
+        .timeout(std::time::Duration::from_secs(3))
+        .build()
+        .ok()?;
+
+    let mut req = client.get(&url);
+    if let Ok(token) = std::env::var("VLLM_API_KEY") {
+        req = req.header("Authorization", format!("Bearer {}", token));
+    }
+    let resp = req.send().ok()?;
+    if !resp.status().is_success() {
+        eprintln!("[openai_compat] vllm {} → {}", url, resp.status());
+        return None;
+    }
+
+    let body: serde_json::Value = resp.json().ok()?;
+    let data = body.get("data")?.as_array()?;
+    let models: Vec<String> = data.iter()
+        .filter_map(|m| m.get("id").and_then(|v| v.as_str()).map(String::from))
+        .collect();
+
+    if models.is_empty() { None } else { Some(models) }
+}
+
 pub async fn stream_run_with_base<F, G>(
     input: RunInput,
     base: String,
@@ -128,7 +163,13 @@ where
         tools: Some(tools_json),
     };
 
-    let engine_name = if base.contains(":1234") || base.contains("lmstudio") { "LM Studio" } else { "Ollama" };
+    let engine_name = if base.contains(":1234") || base.contains("lmstudio") {
+        "LM Studio"
+    } else if base.contains(":8000") || base.contains("vllm") {
+        "vLLM"
+    } else {
+        "Ollama"
+    };
     on_progress(format!("{} ({}) initializing...", engine_name, model));
 
     let client = Client::builder()
@@ -137,10 +178,14 @@ where
         .map_err(|e| AppError::Agent(format!("HTTP client build failed: {}", e)))?;
 
     let mut req = client.post(&url).json(&body);
-    if let Ok(token) = std::env::var("LMSTUDIO_API_KEY") {
-        if engine_name == "LM Studio" {
-            req = req.header("Authorization", format!("Bearer {}", token));
-        }
+    // Apply API key based on engine
+    let api_key = match engine_name {
+        "LM Studio" => std::env::var("LMSTUDIO_API_KEY").ok(),
+        "vLLM" => std::env::var("VLLM_API_KEY").ok(),
+        _ => None,
+    };
+    if let Some(token) = api_key {
+        req = req.header("Authorization", format!("Bearer {}", token));
     }
     let response = req
         .send()
@@ -203,7 +248,13 @@ where
         tools: None,
     };
 
-    let engine_name = if base.contains(":1234") { "LM Studio" } else { "Ollama" };
+    let engine_name = if base.contains(":1234") || base.contains("lmstudio") {
+        "LM Studio"
+    } else if base.contains(":8000") || base.contains("vllm") {
+        "vLLM"
+    } else {
+        "Ollama"
+    };
     on_progress(format!("{} ({}) running (no tools)...", engine_name, model));
 
     let client = Client::builder()
@@ -212,10 +263,13 @@ where
         .map_err(|e| AppError::Agent(format!("HTTP client build failed: {}", e)))?;
 
     let mut req = client.post(&url).json(&body);
-    if let Ok(token) = std::env::var("LMSTUDIO_API_KEY") {
-        if engine_name == "LM Studio" {
-            req = req.header("Authorization", format!("Bearer {}", token));
-        }
+    let api_key = match engine_name {
+        "LM Studio" => std::env::var("LMSTUDIO_API_KEY").ok(),
+        "vLLM" => std::env::var("VLLM_API_KEY").ok(),
+        _ => None,
+    };
+    if let Some(token) = api_key {
+        req = req.header("Authorization", format!("Bearer {}", token));
     }
     let response = req.send().await
         .map_err(|e| AppError::Agent(format!("OpenAI-compatible API 요청 실패: {}", e)))?;
diff --git a/src-tauri/src/commands/agent_detect.rs b/src-tauri/src/commands/agent_detect.rs
index fe3a4b9e..1402a536 100644
--- a/src-tauri/src/commands/agent_detect.rs
+++ b/src-tauri/src/commands/agent_detect.rs
@@ -246,24 +246,82 @@ async fn probe_lmstudio(endpoint: &str) -> AgentDetection {
     det
 }
 
+async fn probe_vllm(endpoint: &str) -> AgentDetection {
+    // vLLM uses OpenAI-compatible /v1/models endpoint
+    let base_raw = endpoint.trim_end_matches('/');
+    let base = if base_raw.ends_with("/v1") { base_raw.to_string() } else { format!("{}/v1", base_raw) };
+    let url = format!("{}/models", base);
+
+    let mut det = AgentDetection {
+        engine: "vllm".into(),
+        kind: "http",
+        installed: false,
+        version: None,
+        path: None,
+        endpoint: Some(base_raw.to_string()),
+        models: vec![],
+        note: None,
+    };
+
+    let client = match reqwest::Client::builder()
+        .timeout(Duration::from_millis(PROBE_TIMEOUT_MS))
+        .build()
+    {
+        Ok(c) => c,
+        Err(e) => { det.note = Some(format!("reqwest build error: {e}")); return det; }
+    };
+
+    eprintln!("[agent-detect] probe vllm: GET {}", url);
+    match client.get(&url).send().await {
+        Ok(resp) if resp.status().is_success() => {
+            match resp.json::<OpenAiModelsResponse>().await {
+                Ok(body) => {
+                    det.installed = true;
+                    det.models = body.data.into_iter().map(|m| m.id).collect();
+                    eprintln!("[agent-detect] vllm ok — {} models", det.models.len());
+                }
+                Err(e) => {
+                    eprintln!("[agent-detect] vllm parse error: {e}");
+                    det.note = Some(format!("응답 파싱 실패: {e}"));
+                }
+            }
+        }
+        Ok(resp) => {
+            let status = resp.status();
+            eprintln!("[agent-detect] vllm status {}", status);
+            det.note = Some(format!("HTTP {status}"));
+        }
+        Err(e) => {
+            eprintln!("[agent-detect] vllm unreachable: {e}");
+            det.note = Some(if e.is_timeout() { "timeout".into() } else { "not reachable".into() });
+        }
+    }
+    det
+}
+
 // ─── Tauri command ───────────────────────────────────────────────────────────
 
 #[tauri::command]
 pub async fn detect_available_agents(
     ollama_endpoint: Option<String>,
     lmstudio_endpoint: Option<String>,
+    vllm_endpoint: Option<String>,
 ) -> Vec<AgentDetection> {
     let ollama_ep = ollama_endpoint.unwrap_or_else(|| "http://localhost:11434".into());
     let lmstudio_ep = lmstudio_endpoint.unwrap_or_else(|| "http://localhost:1234/v1".into());
+    let vllm_ep = vllm_endpoint.unwrap_or_else(|| {
+        std::env::var("VLLM_ENDPOINT").unwrap_or_else(|_| "http://localhost:8000".into())
+    });
 
     // CLI probes — 병렬
-    let (claude, codex, gemini, ollama, lmstudio) = tokio::join!(
+    let (claude, codex, gemini, ollama, lmstudio, vllm) = tokio::join!(
         probe_cli("claude", "claude", &["--version"]),
         probe_cli("codex",  "codex",  &["--version"]),
         probe_cli("gemini", "gemini", &["--version"]),
         probe_ollama(&ollama_ep),
         probe_lmstudio(&lmstudio_ep),
+        probe_vllm(&vllm_ep),
     );
 
-    vec![claude, codex, gemini, ollama, lmstudio]
+    vec![claude, codex, gemini, ollama, lmstudio, vllm]
 }
diff --git a/src-tauri/src/commands/agents.rs b/src-tauri/src/commands/agents.rs
index e8b8167e..cc94b8f2 100644
--- a/src-tauri/src/commands/agents.rs
+++ b/src-tauri/src/commands/agents.rs
@@ -498,9 +498,13 @@ pub async fn start_openai_compat_stream(
 ) -> Result<StartRunResult, AppError> {
     let db = state.inner().clone();
     let db_post = state.inner().clone();
-    let is_lmstudio = input.engine.as_deref() == Some("lmstudio");
-    let engine_label = if is_lmstudio { "lmstudio" } else { "ollama" };
-    eprintln!("[openai-compat] engine={:?} model={:?} is_lmstudio={}", input.engine, input.model, is_lmstudio);
+    let engine_key = input.engine.as_deref().unwrap_or("ollama");
+    let engine_label = match engine_key {
+        "lmstudio" => "lmstudio",
+        "vllm" => "vllm",
+        _ => "ollama",
+    };
+    eprintln!("[openai-compat] engine={:?} model={:?} engine_label={}", input.engine, input.model, engine_label);
     let id_frag = identity_fragment(&input, engine_label);
     let write_arc = db_write_arc(&state);
     let cid = input.conversation_id.clone();
@@ -529,10 +533,10 @@ pub async fn start_openai_compat_stream(
         .filter(|s| !s.is_empty())
         .map(str::to_owned)
         .unwrap_or_else(|| {
-            if is_lmstudio {
-                openai_compat::lmstudio_base_url()
-            } else {
-                std::env::var("OLLAMA_HOST").unwrap_or_else(|_| "http://localhost:11434".into())
+            match engine_label {
+                "lmstudio" => openai_compat::lmstudio_base_url(),
+                "vllm" => openai_compat::vllm_base_url(),
+                _ => std::env::var("OLLAMA_HOST").unwrap_or_else(|_| "http://localhost:11434".into()),
             }
         });
 
@@ -607,7 +611,7 @@ pub fn run_eval_agent(
         "codex" => codex::run(run_input),
         "gemini" => gemini::run(run_input),
         "opencode" => opencode::run(run_input),
-        "ollama" => openai_compat::run(run_input),
+        "ollama" | "vllm" => openai_compat::run(run_input),
         _ => claude::run(run_input),
     };
     let duration_ms = t0.elapsed().as_millis() as i64;
diff --git a/src-tauri/src/commands/model_discovery.rs b/src-tauri/src/commands/model_discovery.rs
index f9bc9b3f..49f64eb2 100644
--- a/src-tauri/src/commands/model_discovery.rs
+++ b/src-tauri/src/commands/model_discovery.rs
@@ -88,6 +88,7 @@ fn fallback_models(engine: &str) -> Vec<(&'static str, &'static str, bool)> {
             ("phi-4:latest", "Phi-4", false),
         ],
         "lmstudio" => vec![],  // LM Studio models are always discovered live
+        "vllm" => vec![],      // vLLM models are always discovered live
         _ => vec![],
     }
 }
@@ -366,7 +367,7 @@ fn discover_lmstudio() -> Option<Vec<String>> {
 
 // ─── Core API ───────────────────────────────────────────────────────────────
 
-const ENGINES: &[&str] = &["claude", "codex", "gemini", "ollama", "lmstudio"];
+const ENGINES: &[&str] = &["claude", "codex", "gemini", "ollama", "lmstudio", "vllm"];
 
 fn get_models_for_engine(engine: &str, force: bool) -> (Vec<String>, String) {
     // Check cache — invalidate early if the tracked binary's mtime has changed
@@ -398,6 +399,7 @@ fn get_models_for_engine(engine: &str, force: bool) -> (Vec<String>, String) {
         },
         "ollama" => (crate::agents::openai_compat::discover_models(), None),
         "lmstudio" => (discover_lmstudio(), None),
+        "vllm" => (crate::agents::openai_compat::discover_vllm(), None),
         _ => (None, None),
     };
 
diff --git a/src-tauri/src/commands/roundtable_helpers/executor.rs b/src-tauri/src/commands/roundtable_helpers/executor.rs
index bb5d2a8e..da176426 100644
--- a/src-tauri/src/commands/roundtable_helpers/executor.rs
+++ b/src-tauri/src/commands/roundtable_helpers/executor.rs
@@ -52,6 +52,7 @@ pub async fn run_participant(
             "gemini" => (gemini::run(run_input), "gemini"),
             "opencode" => (opencode::run(run_input), "opencode"),
             "ollama" => (openai_compat::run(run_input), "ollama"),
+            "vllm" => (openai_compat::run(run_input), "vllm"),
             _ => (
                 Err(AppError::Agent(format!("unsupported engine: {}", engine_key_owned))),
                 "unknown",
@@ -168,7 +169,7 @@ pub(super) async fn stream_participant(
             .await
             .unwrap_or_else(|_| (Err(AppError::Agent("participant task panicked".into())), "unknown"))
         }
-        "ollama" => {
+        "ollama" | "vllm" => {
             let a = app.clone(); let mi = msg_id.clone(); let ci = conversation_id.clone();
             let on_chunk = {
                 let a = a.clone(); let mi = mi.clone(); let ci = ci.clone();
@@ -179,7 +180,8 @@ pub(super) async fn stream_participant(
                 }
             };
             let on_progress = |_: String| {};
-            (openai_compat::stream_run(run_input, on_progress, on_chunk).await, "ollama")
+            let label = if engine_key_owned == "vllm" { "vllm" } else { "ollama" };
+            (openai_compat::stream_run(run_input, on_progress, on_chunk).await, label)
         }
         "opencode" => {
             tokio::task::spawn_blocking(move || {
diff --git a/src/components/tunaflow/AgentAvatar.tsx b/src/components/tunaflow/AgentAvatar.tsx
index 26719acf..1fac4592 100644
--- a/src/components/tunaflow/AgentAvatar.tsx
+++ b/src/components/tunaflow/AgentAvatar.tsx
@@ -11,6 +11,7 @@ const ENGINE_ICONS: Record<string, string> = {
 const ENGINE_INITIALS: Record<string, string> = {
   ollama: "O",
   lmstudio: "L",
+  vllm: "V",
 };
 
 interface AgentAvatarProps {
diff --git a/src/components/tunaflow/CreateRoundtableDialog.tsx b/src/components/tunaflow/CreateRoundtableDialog.tsx
index 4470fc3b..4dffe2bc 100644
--- a/src/components/tunaflow/CreateRoundtableDialog.tsx
+++ b/src/components/tunaflow/CreateRoundtableDialog.tsx
@@ -14,7 +14,7 @@ const RT_MODES: { id: RtMode; label: string; desc: string }[] = [
   { id: "deliberative", label: "Deliberative", desc: "Round 1 independent, Round 2+ reflects on all" },
 ];
 
-const ENGINES = ["claude", "codex", "gemini", "ollama", "lmstudio"] as const;
+const ENGINES = ["claude", "codex", "gemini", "ollama", "lmstudio", "vllm"] as const;
 
 interface CreateRoundtableDialogProps {
   open: boolean;
diff --git a/src/components/tunaflow/MetaAgentSelector.tsx b/src/components/tunaflow/MetaAgentSelector.tsx
index 607222f7..ebbb841c 100644
--- a/src/components/tunaflow/MetaAgentSelector.tsx
+++ b/src/components/tunaflow/MetaAgentSelector.tsx
@@ -46,6 +46,7 @@ const ENGINE_META: Record<string, EngineMeta> = {
   gemini:   { label: "Gemini",    installHintKey: "", installHint: "npm install -g @google/gemini-cli",        docLink: "https://ai.google.dev/gemini-api/docs/cli" },
   ollama:   { label: "Ollama",    installHintKey: "ollama_install_hint",   defaultEndpoint: "http://localhost:11434" },
   lmstudio: { label: "LM Studio", installHintKey: "lmstudio_install_hint", defaultEndpoint: "http://localhost:1234/v1" },
+  vllm:     { label: "vLLM",      installHintKey: "vllm_install_hint",     defaultEndpoint: "http://localhost:8000" },
 };
 
 // CLI engines whose `models` list comes from the dynamic discovery store
@@ -60,6 +61,7 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
   const [detections, setDetections] = useState<AgentDetection[] | null>(null);
   const [ollamaEndpoint, setOllamaEndpoint] = useState("http://localhost:11434");
   const [lmstudioEndpoint, setLmstudioEndpoint] = useState("http://localhost:1234/v1");
+  const [vllmEndpoint, setVllmEndpoint] = useState("http://localhost:8000");
 
   const [selectedEngine, setSelectedEngine] = useState<string | null>(null);
   const [modelByEngine, setModelByEngine] = useState<Record<string, string>>({});
@@ -86,12 +88,13 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
   };
 
   // Initial + on-endpoint-change detection
-  const runDetect = async (oEp: string, lEp: string) => {
+  const runDetect = async (oEp: string, lEp: string, vEp: string) => {
     setDetections(null);
     try {
       const result = await invoke<AgentDetection[]>("detect_available_agents", {
         ollamaEndpoint: oEp,
         lmstudioEndpoint: lEp,
+        vllmEndpoint: vEp,
       });
       setDetections(result);
     } catch (e) {
@@ -107,7 +110,7 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
     if (engineModels.length === 0) {
       loadEngineModels().catch((e) => console.warn("[meta-agent] loadEngineModels", e));
     }
-    runDetect(ollamaEndpoint, lmstudioEndpoint);
+    runDetect(ollamaEndpoint, lmstudioEndpoint, vllmEndpoint);
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, []);
 
@@ -132,15 +135,17 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
     });
   }, [detections, engineModels]);
 
-  const onEndpointChange = (engine: "ollama" | "lmstudio", value: string) => {
+  const onEndpointChange = (engine: "ollama" | "lmstudio" | "vllm", value: string) => {
     if (engine === "ollama") setOllamaEndpoint(value);
-    else setLmstudioEndpoint(value);
+    else if (engine === "lmstudio") setLmstudioEndpoint(value);
+    else setVllmEndpoint(value);
 
     if (debounceRef.current) window.clearTimeout(debounceRef.current);
     debounceRef.current = window.setTimeout(() => {
       const o = engine === "ollama" ? value : ollamaEndpoint;
       const l = engine === "lmstudio" ? value : lmstudioEndpoint;
-      runDetect(o, l);
+      const v = engine === "vllm" ? value : vllmEndpoint;
+      runDetect(o, l, v);
     }, 600);
   };
 
@@ -169,7 +174,11 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
     if (!canProceed || !selectedEngine) return;
     const det = detections!.find((d) => d.engine === selectedEngine)!;
     const endpoint = det.kind === "http"
-      ? (selectedEngine === "ollama" ? ollamaEndpoint : lmstudioEndpoint)
+      ? (selectedEngine === "ollama"
+          ? ollamaEndpoint
+          : selectedEngine === "lmstudio"
+            ? lmstudioEndpoint
+            : vllmEndpoint)
       : undefined;
     onProceed({
       engine: selectedEngine,
@@ -253,8 +262,8 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
                       <span className="text-[10px] text-muted-foreground/60 shrink-0">Endpoint</span>
                       <input
                         type="text"
-                        value={d.engine === "ollama" ? ollamaEndpoint : lmstudioEndpoint}
-                        onChange={(e) => onEndpointChange(d.engine as "ollama" | "lmstudio", e.target.value)}
+                        value={d.engine === "ollama" ? ollamaEndpoint : d.engine === "lmstudio" ? lmstudioEndpoint : vllmEndpoint}
+                        onChange={(e) => onEndpointChange(d.engine as "ollama" | "lmstudio" | "vllm", e.target.value)}
                         className="flex-1 text-[10px] font-mono bg-background border border-border/60 rounded px-2 py-1 focus:outline-none focus:border-primary/60"
                       />
                     </div>
@@ -308,6 +317,8 @@ export function MetaAgentSelector({ onProceed, onSkip, projectName }: Props) {
                             ? t("meta_agent.ollama_install_hint")
                             : meta.installHintKey === "lmstudio_install_hint"
                             ? t("meta_agent.lmstudio_install_hint")
+                            : meta.installHintKey === "vllm_install_hint"
+                            ? t("meta_agent.vllm_install_hint")
                             : meta.installHint}
                         </span>
                       </div>
diff --git a/src/components/tunaflow/input/EngineSelector.tsx b/src/components/tunaflow/input/EngineSelector.tsx
index 6bd63930..e0ea6e4b 100644
--- a/src/components/tunaflow/input/EngineSelector.tsx
+++ b/src/components/tunaflow/input/EngineSelector.tsx
@@ -2,7 +2,7 @@ import { useRef, useEffect, useState } from "react";
 import { cn } from "@/lib/utils";
 import { ChevronDown } from "lucide-react";
 
-type Engine = "claude" | "codex" | "gemini" | "ollama" | "lmstudio";
+type Engine = "claude" | "codex" | "gemini" | "ollama" | "lmstudio" | "vllm";
 
 const ENGINE_LIST: { id: Engine; label: string; color: string }[] = [
   { id: "claude", label: "Claude", color: "text-agent-claude" },
@@ -10,6 +10,7 @@ const ENGINE_LIST: { id: Engine; label: string; color: string }[] = [
   { id: "gemini", label: "Gemini", color: "text-agent-gemini" },
   { id: "ollama", label: "Ollama", color: "text-agent-ollama" },
   { id: "lmstudio", label: "LM Studio", color: "text-agent-lmstudio" },
+  { id: "vllm", label: "vLLM", color: "text-agent-vllm" },
 ];
 
 interface EngineSelectorProps {
diff --git a/src/components/tunaflow/settings/AgentsSection.tsx b/src/components/tunaflow/settings/AgentsSection.tsx
index b84a58e1..56ac5f4d 100644
--- a/src/components/tunaflow/settings/AgentsSection.tsx
+++ b/src/components/tunaflow/settings/AgentsSection.tsx
@@ -19,13 +19,14 @@ import {
 
 // Keep in sync with ENGINE_CONFIGS (src/lib/engineConfig.ts). OpenCode removed;
 // Ollama + LMStudio share the openai-compatible runtime.
-const ENGINES = ["claude", "codex", "gemini", "ollama", "lmstudio"] as const;
+const ENGINES = ["claude", "codex", "gemini", "ollama", "lmstudio", "vllm"] as const;
 const ENGINE_LABELS: Record<(typeof ENGINES)[number], string> = {
   claude: "Claude",
   codex: "Codex",
   gemini: "Gemini",
   ollama: "Ollama",
   lmstudio: "LM Studio",
+  vllm: "vLLM",
 };
 
 export function AgentsSection() {
@@ -75,20 +76,22 @@ export function AgentsSection() {
   // Issue #175 MVP — Ollama / LM Studio base URL override.
   // Keyed by engine; appStore key pattern `engineEndpoint:{engine}` mirrors
   // existing conventions (activeSkills:{pk}, skillDetectionDismissed:{pk}).
-  const [endpointOverride, setEndpointOverride] = useState<Record<"ollama" | "lmstudio", string>>({
+  const [endpointOverride, setEndpointOverride] = useState<Record<"ollama" | "lmstudio" | "vllm", string>>({
     ollama: "",
     lmstudio: "",
+    vllm: "",
   });
   useEffect(() => {
     (async () => {
-      const [ol, ls] = await Promise.all([
+      const [ol, ls, vl] = await Promise.all([
         getSetting<string>("engineEndpoint:ollama", ""),
         getSetting<string>("engineEndpoint:lmstudio", ""),
+        getSetting<string>("engineEndpoint:vllm", ""),
       ]);
-      setEndpointOverride({ ollama: ol, lmstudio: ls });
+      setEndpointOverride({ ollama: ol, lmstudio: ls, vllm: vl });
     })();
   }, []);
-  const handleEndpointChange = async (engine: "ollama" | "lmstudio", value: string) => {
+  const handleEndpointChange = async (engine: "ollama" | "lmstudio" | "vllm", value: string) => {
     const trimmed = value.trim();
     setEndpointOverride((prev) => ({ ...prev, [engine]: trimmed }));
     await setSetting(`engineEndpoint:${engine}`, trimmed);
@@ -160,20 +163,22 @@ export function AgentsSection() {
               </div>
             </div>
 
-            {/* Issue #175 MVP — Ollama / LM Studio base URL override */}
-            {(selected.engine === "ollama" || selected.engine === "lmstudio") && (
+            {/* Issue #175 MVP — Ollama / LM Studio / vLLM base URL override */}
+            {(selected.engine === "ollama" || selected.engine === "lmstudio" || selected.engine === "vllm") && (
               <div>
                 <label className="text-tf-sm text-muted-foreground mb-1 block">
                   {t("agents.endpoint.label")}
                 </label>
                 <input
                   type="text"
-                  value={endpointOverride[selected.engine]}
-                  onChange={(e) => handleEndpointChange(selected.engine as "ollama" | "lmstudio", e.target.value)}
+                  value={endpointOverride[selected.engine as "ollama" | "lmstudio" | "vllm"]}
+                  onChange={(e) => handleEndpointChange(selected.engine as "ollama" | "lmstudio" | "vllm", e.target.value)}
                   placeholder={
                     selected.engine === "ollama"
                       ? t("agents.endpoint.placeholder_ollama")
-                      : t("agents.endpoint.placeholder_lmstudio")
+                      : selected.engine === "lmstudio"
+                        ? t("agents.endpoint.placeholder_lmstudio")
+                        : "http://localhost:8000"
                   }
                   className="w-full bg-background rounded-lg px-3 py-2 text-tf-caption font-mono outline-none border border-border/30 focus:border-ring/40"
                 />
diff --git a/src/components/tunaflow/settings/RuntimeSection.tsx b/src/components/tunaflow/settings/RuntimeSection.tsx
index d3fb3db4..9498937a 100644
--- a/src/components/tunaflow/settings/RuntimeSection.tsx
+++ b/src/components/tunaflow/settings/RuntimeSection.tsx
@@ -301,6 +301,7 @@ function InsightAgentConfig() {
             <option value="gemini">Gemini</option>
             <option value="ollama">Ollama</option>
             <option value="lmstudio">LM Studio</option>
+            <option value="vllm">vLLM</option>
           </select>
         </div>
         <div className="space-y-1 flex-1">
diff --git a/src/index.css b/src/index.css
index 7ccf2f49..46d63983 100644
--- a/src/index.css
+++ b/src/index.css
@@ -32,6 +32,7 @@
   --agent-opencode: oklch(0.62 0.18 145);
   --agent-ollama: oklch(0.65 0.15 30);
   --agent-lmstudio: oklch(0.60 0.16 180);
+  --agent-vllm: oklch(0.58 0.20 150);
 
   --status-draft: oklch(0.65 0.15 55);
   --status-approved: oklch(0.62 0.18 145);
@@ -119,6 +120,7 @@
   --color-agent-opencode: var(--agent-opencode);
   --color-agent-ollama: var(--agent-ollama);
   --color-agent-lmstudio: var(--agent-lmstudio);
+  --color-agent-vllm: var(--agent-vllm);
   --color-status-draft: var(--status-draft);
   --color-status-approved: var(--status-approved);
   --color-status-rejected: var(--status-rejected);
diff --git a/src/lib/engineConfig.ts b/src/lib/engineConfig.ts
index 3082a31d..6b5c09e8 100644
--- a/src/lib/engineConfig.ts
+++ b/src/lib/engineConfig.ts
@@ -11,4 +11,5 @@ export const ENGINE_CONFIGS: Record<string, EngineConfig> = {
   gemini:   { command: "start_gemini_stream",        engineKey: "gemini",      label: "Gemini initializing...",   hasChunkEvent: true },
   ollama:   { command: "start_openai_compat_stream", engineKey: "ollama",      label: "Ollama initializing...",   hasChunkEvent: true },
   lmstudio: { command: "start_openai_compat_stream", engineKey: "lmstudio",    label: "LM Studio initializing...", hasChunkEvent: true },
+  vllm:     { command: "start_openai_compat_stream", engineKey: "vllm",        label: "vLLM initializing...",      hasChunkEvent: true },
 };
diff --git a/src/lib/initialSetupApply.ts b/src/lib/initialSetupApply.ts
index 8965e50d..5f3521c0 100644
--- a/src/lib/initialSetupApply.ts
+++ b/src/lib/initialSetupApply.ts
@@ -22,7 +22,7 @@ import type { AgentProfile } from "@/types";
 const KNOWN_PERSONAS = new Set(DEFAULT_PERSONAS.map((p) => p.id));
 
 /** Engines we recognize. Matches ENGINE_CONFIGS keys. */
-const KNOWN_ENGINES = new Set(["claude", "codex", "gemini", "ollama", "lmstudio", "openai"]);
+const KNOWN_ENGINES = new Set(["claude", "codex", "gemini", "ollama", "lmstudio", "vllm", "openai"]);
 
 export interface RecommendedProfile {
   role: string;        // "architect" | "developer" | "reviewer" | …
diff --git a/src/lib/sendPipeline/buildSendInput.ts b/src/lib/sendPipeline/buildSendInput.ts
index 800ef255..e6678622 100644
--- a/src/lib/sendPipeline/buildSendInput.ts
+++ b/src/lib/sendPipeline/buildSendInput.ts
@@ -49,7 +49,7 @@ export async function buildSendInput(p: BuildSendInputParams): Promise<SendWithC
   // Empty string is equivalent to "no override" so the backend falls back to
   // env var / hardcoded default; only non-empty trimmed value is forwarded.
   let customBaseUrl: string | undefined;
-  if (p.engine === "ollama" || p.engine === "lmstudio") {
+  if (p.engine === "ollama" || p.engine === "lmstudio" || p.engine === "vllm") {
     const raw = await getSetting<string>(`engineEndpoint:${p.engine}`, "");
     const trimmed = raw.trim();
     if (trimmed) customBaseUrl = trimmed;
diff --git a/src/locales/en/dialog.json b/src/locales/en/dialog.json
index bbe28186..ff99571b 100644
--- a/src/locales/en/dialog.json
+++ b/src/locales/en/dialog.json
@@ -99,6 +99,7 @@
     "skip_overlay_confirm": "Skip and proceed",
     "ollama_install_hint": "Install from https://ollama.com/download then run `ollama serve`",
     "lmstudio_install_hint": "Install from https://lmstudio.ai then start the Local Server",
+    "vllm_install_hint": "Install with `pip install vllm` then run `vllm serve <model> --host 0.0.0.0 --port 8000`",
     "model_loading": "Loading models...",
     "model_empty": "No models — install or refresh via Settings → Runtime"
   },
diff --git a/src/locales/ko/dialog.json b/src/locales/ko/dialog.json
index a24b84a3..16e3e20f 100644
--- a/src/locales/ko/dialog.json
+++ b/src/locales/ko/dialog.json
@@ -99,6 +99,7 @@
     "skip_overlay_confirm": "건너뛰고 진행",
     "ollama_install_hint": "https://ollama.com/download 에서 설치 후 `ollama serve`",
     "lmstudio_install_hint": "https://lmstudio.ai 에서 설치 후 Local Server 시작",
+    "vllm_install_hint": "`pip install vllm` 후 `vllm serve <model> --host 0.0.0.0 --port 8000`",
     "model_loading": "모델 목록 로딩 중...",
     "model_empty": "모델 없음 — 설치 또는 Settings → Runtime 에서 새로고침"
   },
diff --git a/src/types/index.ts b/src/types/index.ts
index ad3ede8a..756485b1 100644
--- a/src/types/index.ts
+++ b/src/types/index.ts
@@ -182,7 +182,7 @@ export interface SendWithClaudeInput {
   userMessageId?: string;
   prompt: string;
   model?: string;
-  /** Engine key for backend routing (e.g. "ollama" vs "lmstudio") */
+  /** Engine key for backend routing (e.g. "ollama" / "lmstudio" / "vllm") */
   engine?: string;
   /** Passed directly when no agent is selected */
   systemPrompt?: string;
@@ -205,16 +205,16 @@ export interface SendWithClaudeInput {
   /** Absolute paths of image attachments — used by Codex CLI (`-i <path>`).
    *  Other engines read the image via `Read` tool from the prompt path section. */
   imagePaths?: string[];
-  /** Base URL override for OpenAI-compatible engines (ollama / lmstudio).
-   *  Empty/undefined falls back to env var (OLLAMA_HOST / LMSTUDIO_ENDPOINT)
-   *  then the hardcoded default. Issue #175 MVP. */
+  /** Base URL override for OpenAI-compatible engines (ollama / lmstudio / vllm).
+   *  Empty/undefined falls back to env var (OLLAMA_HOST / LMSTUDIO_ENDPOINT /
+   *  VLLM_ENDPOINT) then the hardcoded default. Issue #175 MVP. */
   customBaseUrl?: string;
 }
 
 export interface RoundtableParticipant {
   name: string;
   model?: string;
-  /** "claude" | "codex" | "gemini" | "ollama" | "lmstudio" — defaults to "claude" on backend */
+  /** "claude" | "codex" | "gemini" | "ollama" | "lmstudio" | "vllm" — defaults to "claude" on backend */
   engine?: string;
   /** Blind verifier — receives topic only, no prior/current transcript */
   blind?: boolean;