Use the MCP server

This guide shows you how to use Control Center’s MCP server with external clients.

What is MCP?

The Model Context Protocol (MCP) is a JSON-RPC 2.0 protocol for exposing tools to AI clients. Control Center runs an MCP server that exposes ~70 typed tools for agents and external clients.

Start the MCP server

The MCP server starts automatically when Control Center runs. It listens for JSON-RPC connections from local clients.

Available tools

Tools are organized by domain:

Domain	Tools	Description
Agents	list, hire, update, fire, kill, run logs	Agent lifecycle management
Agent collaboration	consult_agent, propose_hire, request_peer_review	Agent-to-agent coordination
Skills	list, create	Skill discovery and authoring
Workspaces & repos	list/create workspaces, list_repos	Workspace and repository management
Messaging	channels, channels, messages, thread replies	Communication
PR & review	start_ai_review, review nodes, verdicts, dispatch_reviewers, finalize_review, publish_review_to_github	PR review workflow
Memory	facts, policies, domains, observations, search_memory, working-memory notes	Knowledge management
Code graph	search_code, code_symbol, code_callers, code_callees, code_impact	Code navigation
Resources	read	File/resource reading
User interaction	ask_user_question, request_confirmation, suggest_tasks	Human-in-the-loop
Doctor	doctor	Environment diagnostics
Ticketing & projects	CRUD, lifecycle, assignment, links, approval gates, projects	Work tracking

See MCP tools reference for the complete catalog.

The calendar and meetings features are desktop-only and are not exposed over MCP.

Workspace scoping

Every tool that touches workspace-scoped data requires workspace_id in its parameters:

{
  "workspace_id": "abc-123",
  "title": "Fix login bug",
  "priority": "high"
}

Tools that span all workspaces (e.g. list_workspaces, create_workspace) are the only exceptions.

Mode gating

The ModeToolGuard restricts which tools are available based on the current mode. This prevents agents in plan mode from using write tools, for example.

Connect from a client

Configure your MCP client to connect to Control Center’s server endpoint. The exact configuration depends on your client.

Destructive action confirmation

Some tools require explicit confirmation before execution:

The tool returns an ApprovalPayload describing the action
The client must confirm before the tool proceeds
This provides a safety net for destructive operations