ACP Governance MCP Server
Model Context Protocol server that lets Claude, ChatGPT, Cursor, Lovable, and any MCP client check tool calls against Agentic Control Plane governance — policy decisions, rate limits, audit logs, identity attribution.
One sentence: before your AI agent runs a sensitive tool, it asks ACP whether the call is allowed. ACP says yes, no, or asks for confirmation, and writes an audit row attributable to the human behind the agent.
What it exposes
Two tools, callable via MCP:
| Tool | What it does |
|---|---|
acp_check | Answer "would this call be allowed?" for a specific call — for explicit questions or pre-flight planning. If ACP hooks are installed in the harness, tool calls are already governed automatically and this doesn't need to be called per tool. Returns allow / deny / ask plus a reason. |
acp_status | Verify the connection and your workspace identity. |
That's the whole surface. Everything else — policies, audit logs, scope intersection, delegation chains — runs server-side at api.agenticcontrolplane.com. This MCP server is just the bridge.
Install (hosted — recommended)
The server is hosted at https://mcp.agenticcontrolplane.com/mcp. Add it as a connector in your MCP client:
Claude Desktop / Claude.ai connector:
URL: https://mcp.agenticcontrolplane.com/mcp
Auth: OAuth (sign in with Google through ACP)
ChatGPT, Cursor, Lovable, Cline: Same URL, same OAuth flow. Most clients have a one-click "Add MCP Server" UI.
Programmatic clients (your own agent code):
POST https://mcp.agenticcontrolplane.com/mcp
Authorization: Bearer gsk_<your-acp-api-key>
Content-Type: application/json
You'll need an ACP workspace. The free tier is unlimited tool-call logging — sign up at cloud.agenticcontrolplane.com/login.
Install (self-host)
If you want to run the bridge yourself — for air-gapped deployments, or to point at a self-hosted ACP gateway — clone and run:
git clone https://github.com/davidcrowe/acp-mcp-server
cd acp-mcp-server
npm install
npm run build
# Point at the ACP API (default: https://api.agenticcontrolplane.com)
export ACP_API_BASE=https://your-acp-gateway.example.com
# Optional: service-level API key for OAuth users (ChatGPT, Claude.ai)
# whose JWTs aren't directly usable as ACP tokens
export ACP_SERVICE_KEY=gsk_workspace_...
npm start
# → MCP endpoint: POST http://0.0.0.0:3000/mcp
# → OAuth discovery: GET /.well-known/oauth-protected-resource
The bridge speaks streamable-HTTP MCP and proxies to ACP's /govern/tool-use endpoint.
How it fits
your AI client this MCP server ACP gateway
───────────── ───────────────── ───────────
Claude / ChatGPT ──► mcp.agenticcontrolplane ──► api.agenticcontrolplane
Cursor / Lovable POST /mcp (acp_check) POST /govern/tool-use
↓
policy + audit + identity
↓
allow / deny / ask
Every call writes an audit row attributable to the human identity behind the OAuth session — so you get a complete log of every governed tool call across every MCP client your team uses, in one workspace.
Auth model
The server supports two authentication paths:
- OAuth (recommended for human-driven clients) — Claude.ai, ChatGPT, etc. complete an OAuth flow against ACP's identity provider; the resulting Auth0 JWT identifies the human. The bridge uses an
ACP_SERVICE_KEYto authorize the underlying governance call on the human's behalf. - Bearer
gsk_API key (recommended for programmatic clients) — pass an ACP API key directly asAuthorization: Bearer gsk_.... The key's identity is the ACP-side actor. Skip OAuth.
OAuth discovery metadata is served at /.well-known/oauth-protected-resource per the MCP authorization spec.
Local development
npm install
npm run dev # tsx-based hot reload
Tools are defined in src/tools/tools.ts. The MCP JSON-RPC handler is in src/handlers/mcpHandler.ts. The Express entry point and rate limits are in src/server/expressServer.ts.
License
MIT — see LICENSE.