Acp MCP server

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:

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:

1. 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 ACPSERVICEKEY to authorize the underlying governance call on the human's behalf.
2. Bearer gsk API key (recommended for programmatic clients) — pass an ACP API key directly as Authorization: Bearer gsk.... The key's identity is the ACP-side actor. Skip OAuth.

Shortened. The full README is on GitHub.