mycelium

A coordination layer for multi-agent systems: shared rooms, persistent memory, and mediated negotiation over a SLIM node.

---

mycelium-promo.mp4

install → coordinate → plan → work.

---

The Problem

Very little exists for agents operating as autonomous peers on a shared mission. To get reliable results, practitioners reach for an orchestrator, a predefined workflow, or a tightly defined handoff structure. Users attempting peer agent coordination have to manually construct scaffolding for memory sharing and context passing. And even then, without coordination infrastructure, the result is AI theatre: agents that talk over each other, repeat work already done, fail to recognise disagreement, and fail to negotiate trade-offs.

Who Mycelium Is For

Mycelium is built for autonomous agents operating as peers, with no predefined workflow, no centralized supervisor, and no hierarchy. That includes agents like Claude Code: given a mission and a tool allowlist, left to plan and execute without step-by-step human approval.

Alignment pays off at 3+ agents. At three it improves decision quality over uncoordinated approaches; at four or more it's often the difference between converging on a shared answer and not converging at all.

If your system has a central orchestrator routing tasks to worker agents, you probably don't need Mycelium: your orchestrator is already the coordination layer. Mycelium is for the case where there is no orchestrator, and you don't want one.

Does It Work

Mycelium was evaluated across 14 decision scenarios in a controlled A/B study. See Evaluation Results for the full findings.

What Mycelium Does

Mycelium provides coordination functions for autonomous agents operating as peers. The first: alignment, agreeing on a shared position at the start of a mission or any point during it, so decisions don't get re-litigated, work doesn't get duplicated, and every agent that joins inherits what the others already know.

Mycelium gives agents rooms to coordinate in, persistent memory that accumulates within a room, and an aligner that mediates negotiation so every agent has a voice and the team arrives at a single shared answer.

Two surfaces, one room, built for each other. You and your agents coordinate together:

• You work in the UI: create a room, add agents, hand them a mission, and watch them reach a shared decision and a plan, live.
• Your agents work through the CLI: they join the room, negotiate, and write to shared memory on their own (that's what the mycelium skill teaches them).

That's also why you need at least one agent runtime (Claude Code): the agents aren't an optional add-on, they're half the system.

# Agent 1 shares context in a persistent room
mycelium memory set "position/julia" "I think we should use REST, not GraphQL" --handle julia-agent

# Agent 2 (hours later, different session) reads and adds their perspective
mycelium memory search "API design decisions"
mycelium memory set "position/selina" "Agree on REST, but we need pagination standards" --handle selina-agent

When agents need to agree on something, one participant summons the aligner, and each agent takes turns responding until the team converges:

# Register the mediator once, then summon it on the open question
mycelium engine create aligner --kind aligner --room design
mycelium engine invoke aligner "converge on API design"

# Each participant loops: wait for its turn, then post a position
mycelium await   --room design --handle julia-agent --json
mycelium respond --room design --handle julia-agent "I can accept REST with pagination standards."

# On agreement the agreement is compiled into the room's shared plan
mycelium plan tasks   # the - [ ] checklist the team now executes against

How It Works

1. Alignment. When agents need to agree, one participant summons the aligner, a first-party mediator that runs a real NEGMAS Stacked Alternating Offers negotiation. It discovers the issues from the agents' opening positions, brokers each round, addresses one agent at a time, interprets each reply, and stops the instant the agents agree. Every agent has a voice, and the result is one shared answer, not parallel outputs a human has to reconcile. From that consensus Mycelium compiles a shared plan: a - [ ] checklist at plan/tasks.md with @handle owners the whole team executes against. The arc is one line: summon → negotiate → plan → work. The negotiation decides what; the plan is how the team carries it out.

2. Room Memory. Rooms are folders. Memories are markdown files at ~/.mycelium/rooms/{room}/{namespace}/{key}.md. Any agent with file I/O can read and write room memory directly. The CLI is sugar. Memories accumulate across agents and turns, and are searchable by meaning via a local embedding index, with no external service and no database.

3. Peer Collaboration Environment. Any agent joining a room reads from ~/.mycelium/rooms/{room}/ and instantly inherits everything the swarm has learned: decisions made, what failed, open questions, the room's shared plan. No repeated context-setting. Intelligence compounds instead of resetting.

Quick Start

You'll need Docker, an LLM API key (agents can't negotiate without one), and at least one agent runtime (Claude Code).

Onboard your agent. The fastest setup is to let an agent do it — paste this prompt into Claude Code (or any agent runtime with a shell):

Use curl to read https://mycelium-io.github.io/mycelium/agents.md and perform the setup to install Mycelium

The agent follows agents.md, a setup runbook written for agents: it installs the CLI, brings up the stack, and connects its own runtime as an adapter.

Or install by hand:

# 1. Install the CLI and bring up the stack
curl -fsSL https://mycelium-io.github.io/mycelium/install.sh | bash
mycelium install      # pulls images, prompts for your LLM key, writes ~/.mycelium/config.toml

# 2. Open the app: this is where you work
mycelium ui open

From the UI you:

1. create a room (a shared space for agents, memory, and the plan),
2. add agents to it (one per role),
3. give them a mission in the chat box and @mention them,
4. watch them negotiate live to a single shared answer that compiles into the room's plan.

Your agents drive that same room from the CLI on their own, waiting for their turn, responding, and writing to shared memory (that's what the mycelium skill teaches them). You don't run those by hand; they do.

Prefer to script the human side too? Every UI action has a CLI equivalent:

mycelium room create my-project && mycelium room use my-project
mycelium engine create aligner --kind aligner --room my-project
mycelium agent create planner --adapter claude-code --description "Sprint planner"
mycelium engine invoke aligner "converge on the Q3 migration plan"
mycelium plan tasks   # the shared - [ ] checklist the team executes against

Architecture

Memories live on the filesystem. Rooms are folders, memories are markdown files with YAML frontmatter at ~/.mycelium/rooms/{room}/{key}.md. This is the source of truth. Direct writes (cat, editor, agent file I/O) always work; run mycelium memory reindex to refresh the search index after bypassing the CLI. Search runs against a local embedding index (~384-dim, on-device), with no external vector service and no database.

One SLIM node coordinates the room. Agents coordinate over an AGNTCY SLIM group channel per room: MLS-encrypted, shared-secret PSK auth. An always-on thin FastAPI backend is each room's moderator; the agents (and you, by proxy) are members. There's no database, no message broker, no separate realtime service.

Participation is a CLI primitive. Any already-awake caller joins a room and coordinates with two stateless calls: mycelium await long-polls until a message is addressed to its handle (the backend holds membership via a presence lease and a durable transcript cursor, so a tick is never missed between turns), and mycelium respond posts a reply or position. No background process required. An optional daemon exists to auto-wake runtimes that can't wake themselves.

Rooms are git-friendly. Commit ~/.mycelium/rooms/ to share context across machines. Agents on different machines pull the folder and inherit the room's full memory.

Mycelium speaks IOC L9. Coordination rides SLIM as additive Layer 9 epistemic envelopes (exchange for ticks/replies, commit:converged|resolved|rejected, knowledge) with episodes and causal message threading. Summoning the aligner opens an episode: a tagged, membership-scoped negotiation on the room's channel with its own record at log/episodes/{id}.md (the full causally-linked envelope chain), surfaced live in the UI protocol inspector. Agents can state confidence, cite evidence, and flag deference on replies; consensus gets measurable quality metrics. All of it is optional; agents never need to speak L9.

Deployment modes. By default everything runs on a single device (your laptop): backend, SLIM node, agents, and CLI all on localhost. That's the primary target and what mycelium install sets up out of the box. For small teams that want to share memory and coordination state, Mycelium supports a hub-and-spoke mode: one machine runs mycelium hub host to stand up the SLIM node and prints its address; teammates run mycelium connect http://<hub-ip>:<port> to point their CLI + agents at it. mycelium doctor auto-detects which mode you're in.

Room folders use standard namespaces:

~/.mycelium/rooms/{room}/
├── plan/         Shared checklist compiled from negotiation consensus
├── decisions/    Why choices were made
├── status/       Current state of things
├── context/      Background & constraints
├── work/         In-progress and completed work
├── procedures/   How-to guides and runbooks
└── log/          Events, observations, and episode records

Repo layout:

.mycelium/            Memory storage (rooms are folders, memories are markdown files)
mycelium-cli/         CLI + adapters
fastapi-backend/      FastAPI moderator + aligner
mycelium-client/      Generated typed OpenAPI client
mycelium-frontend/    Next.js UI

Adapters

Mycelium reaches your agents through per-runtime adapters. Support is honest about maturity:

Adapter	Status
claude_code	✅ proven
cursor	⚠️ untested / unverified

Claude Code. Installs the mycelium skill (~/.claude/skills/mycelium/SKILL.md), giving Claude Code memory and coordination commands via /mycelium. This is the proven path.

mycelium adapter add claude-code

Development

cd fastapi-backend
uv sync --group dev
uv run pytest tests/ -x -q
uv run ruff check . && uv run ruff format . && uv run ty check .

Interactive API docs at http://localhost:8000/docs when the backend is running.

Built On

Mycelium builds on OSS projects we found invaluable in this space:

• AGNTCY SLIM: the encrypted group-messaging transport agents coordinate over
• IOC L9 protocol models: the epistemic envelope layer that rides SLIM
• NegMAS: multi-issue negotiation, the aligner's engine
• FastAPI + fastembed: the moderator backend and on-device embeddings