MCP server

Use the Council from Claude Desktop, Cursor, Cline, or any MCP-aware client.

Install

pip install mcp>=1.0

The mcp package is required for the FastMCP server runtime. It is NOT a direct dependency of agent-council itself — only the MCP server needs it.

Configure your client

Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "agent-council": {
      "command": "python",
      "args": ["<absolute-path>/agent-council/mcp/agent_council_mcp_server.py"],
      "env": {
        "COUNCIL_CONFIG": "<absolute-path>/agent-council/council.yaml"
      }
    }
  }
}

Restart Claude Desktop.

Cursor

Settings → MCP → Add new server. Same shape as above.

Cline / Roo Cline (VS Code)

Edit cline_mcp_settings.json via the Cline extension. Same shape.

Custom MCP client

The server uses stdio transport by default. Spawn the process, send JSON-RPC requests, read responses on stdout. See the MCP spec for the protocol.

Tools

Tool	Purpose	Status
`council_review`	Run the 5-deliberator Council on a single artifact	✅ v0.1.1
`council_sweep`	Bulk review over watch paths	🔜 v0.1.2
`council_audit`	Query the append-only log	🔜 v0.1.2

Verify

After registering, in Claude Desktop (or your client) ask:

“List the agent-council tools and call council_review on path/to/artifact.md.”

You should see the tool registered and receive a full verdict structure back.

Manual test

python mcp/agent_council_mcp_server.py

The server starts and waits on stdio. Press Ctrl+C to exit.

Limitations

council.yaml must be resolvable — either via the COUNCIL_CONFIG env var, the config argument on each tool call, or by running from a directory containing council.yaml.
Real Council runs are not fast — council_review on a 1,500-word artifact via claude_cli takes 5–7 minutes. Set generous client timeouts.
For CI / smoke testing, point council.yaml#runtime.type at mock_cli so verdicts come back instantly with canned responses.