Pull your OpenClaw agent into any MCP-compatible tool — Claude Code, Codex, Cursor, or anything that speaks MCP.
Your agent already has a persona, memory, and context. Collab MCP makes that available everywhere you work.
Collab MCP is a lightweight server that reads your agent's workspace and exposes three tools via MCP:
| Tool | What it does |
|---|---|
| agent-bootstrap | Loads the agent's identity, memory, stances, and current state into a new session |
| recall | Searches the agent's memory and notes for context on a specific topic |
| record | Writes observations, decisions, and lessons back into the agent's memory |
When you invoke your agent in Claude Code (or any MCP client), it arrives with everything it knows — your goals, recent decisions, open questions, and its own opinions. It doesn't start from scratch.
An OpenClaw agent's workspace is a folder of markdown files:
~/.openclaw/workspace/
├── SOUL.md # Who the agent is
├── USER.md # Who it's working with
├── IDENTITY.md # Name, emoji, metadata
├── MEMORY.md # Long-term memory
├── memory/ # Daily notes (YYYY-MM-DD.md)
├── notes/ # Check-in notes
└── initiative/ # Initiative pipeline output
└── output/
└── YYYY-MM-DD/
└── initiative_brief.md
Collab MCP reads these files and serves them over MCP. That's it. No database, no cloud sync, no account required. Your agent's workspace is the source of truth.
If your agent uses Seedvault for note sync, Collab MCP can also search those notes via the recall tool.
If you're running an OpenClaw agent, give it the setup skill:
Read https://raw.githubusercontent.com/collaborator-ai/collab-mcp/main/SKILL.md and follow it.
Your agent will clone the repo, install dependencies, configure the server, and connect it to your coding tools. You can review the skill file first — it's short and readable.
# 1. Clone and build
git clone https://github.com/collaborator-ai/collab-mcp.git
cd collab-mcp
npm install
npm run build
# 2. Set environment variables
export OPENCLAW_WORKSPACE=~/.openclaw/workspace # Path to your agent's workspace
export OPENCLAW_AUTH_TOKEN=$(openssl rand -hex 16) # API key for connections
export PORT=3100
# 3. Start the server
node build/index.js
# 4. Connect to Claude Code
claude mcp add collab --transport sse \
"http://localhost:3100/mcp/sse?key=$OPENCLAW_AUTH_TOKEN"That's it. Open Claude Code and your agent's context is available.
codex mcp add collab --url "http://localhost:3100/mcp?key=$OPENCLAW_AUTH_TOKEN"For access from other machines, put a tunnel in front:
cloudflared tunnel --url http://localhost:3100Then set PUBLIC_URL to the tunnel URL and use that in your claude mcp add command.
| Variable | Required | Description |
|---|---|---|
OPENCLAW_WORKSPACE |
Yes | Path to the agent's workspace |
OPENCLAW_AUTH_TOKEN |
Recommended | API key for MCP connections. Without it, the server is open. |
PORT |
No | Server port (default: 3100) |
TRANSPORT |
No | sse or stdio (default: sse) |
PUBLIC_URL |
No | Public URL if behind a tunnel |
SV_SERVER |
No | Seedvault server URL |
SV_TOKEN |
No | Seedvault API token |
SV_CONTRIBUTOR |
No | Seedvault contributor to search |
When a coding tool calls agent-bootstrap, it receives:
- Persona — from
SOUL.md(who the agent is, how it thinks) - Operating instructions — from
AGENTS.md - User profile — from
USER.md - Current state — from the latest initiative brief (what's active, what's stuck)
- Active stances — opinions and decisions extracted from memory
- Recent context — from daily notes and check-ins
The agent doesn't recite this back. It internalizes it, orients in the current repo, and proposes what to do based on everything it knows.
- The server only reads from your local workspace. Nothing is sent to external services unless you configure Seedvault.
- The auth token is a simple API key passed as a URL parameter. This is fine for local use. For remote access, always use HTTPS (e.g., Cloudflare Tunnel).
- OAuth 2.1 is also supported for clients that need it (e.g., Claude Desktop). The flow auto-approves — no login page.
MIT