Quickstart

You'll start the capability server and point an MCP client at it. Assumes the CLI and Camofox are installed — see Install.

1. Start Camofox + the daemon

bureau start

bureau start launches Camofox (:9377) and then bureau serve, which hosts the MCP catalogue on port 8830. Useful flags:

• --port N — daemon port (default BUREAU_PORT / PORT env, else 8830).
• --camofox-port N — Camofox port (default 9377).
• --detach — spawn the daemon in the background and exit once it's healthy.
• bureau start bureau-only — skip Camofox (you're already running it on :9377).

Check it's up:

curl http://localhost:8830/health
# {"ok":true,"tools":<n>}

2. Connect from Claude Code

Bureau exposes a stateless Streamable HTTP MCP endpoint at POST http://localhost:8830/mcp. Add it to your MCP client's config — for Claude Code, an .mcp.json:

{
  "mcpServers": {
    "bureau": {
      "type": "http",
      "url": "http://localhost:8830/mcp"
    }
  }
}

The same URL works for any client that speaks the Streamable HTTP transport (Cursor, agentproto, …). For a stdio-only client, bridge it with an HTTP-to-stdio adapter such as mcp-remote.

3. Drive the browser

Your agent now sees Bureau's tools. Try an anonymous read:

• browser_navigate to open a URL, then scrape to pull readable text, or
• scrape with a url directly.

To act as one of your logged-in identities, first save a session (Sessions) and pass its id to scrape / browser_act / the social_* tools.

4. (Optional) AI actions

Tools like research_run and the llm.ask / intel.judge reasoning seam need a model. Either set ANTHROPIC_API_KEY (your own key — free, offline) or link a Bureau account for the metered proxy. See Account & AI credits.

Endpoints reference