agent-mesh-mcp

MCP server for inter-agent communication via RabbitMQ. Two transports:

• stdio (index.mjs) — for local Claude Code / Cursor sessions

• HTTP (server.mjs) — for remote access from Claude.ai, Claude Desktop, or mobile via Streamable HTTP + Keycloak OAuth

Install

npm install -g agent-mesh-mcp

Related MCP server: gptqueue

Local Usage (stdio)

Claude Code (~/.claude/settings.local.json)

{
  "mcpServers": {
    "agent-mesh": {
      "command": "agent-mesh-mcp",
      "env": {
        "RABBITMQ_URL": "amqp://user:pass@your-rabbitmq:5672/",
        "RABBITMQ_MGMT_URL": "http://your-rabbitmq:15672",
        "EXCHANGE_NAME": "agents",
        "AGENT_NAME": "human"
      }
    }
  }
}

Cursor (.cursor/mcp.json)

Same format as above.

Remote Usage (HTTP)

Run server.mjs as a standalone HTTP server with Keycloak OAuth:

MCP_PORT=3100 \
KEYCLOAK_URL=https://your-keycloak \
KEYCLOAK_REALM=master \
KEYCLOAK_CLIENT_ID=agent-mesh \
KEYCLOAK_CLIENT_SECRET=your-secret \
RABBITMQ_URL=amqp://user:pass@your-rabbitmq:5672/ \
RABBITMQ_MGMT_URL=http://your-rabbitmq:15672 \
EXCHANGE_NAME=agents \
node server.mjs

Then add to Claude Code (.mcp.json or .claude.json):

{
  "mcpServers": {
    "agent-mesh": {
      "type": "http",
      "url": "https://mesh.openbiocure.ai/mcp"
    }
  }
}

Claude Code auto-discovers OAuth via /.well-known/oauth-authorization-server and handles the Keycloak login flow in your browser.

Architecture

Claude.ai / Phone / Desktop
        |
        v  HTTPS
   your-domain/mcp (Cloudflare Tunnel or reverse proxy)
        |
        v
   server.mjs (Streamable HTTP + OAuth)
        |
        v  AMQP
   RabbitMQ --> All workers (any machine)

OAuth Flow

The server proxies OAuth to Keycloak:

• GET /authorize — redirects to Keycloak login

• POST /token — exchanges auth code for access token via Keycloak

• POST /register — dynamic client registration

• GET /.well-known/oauth-authorization-server — OAuth metadata

• GET /health — health check (no auth)

All /mcp requests require a valid Keycloak bearer token.

Tools

list_agents()

List all online agents and their topics.

ask_agent(topic, message, timeout?)

Ask another agent a question via RabbitMQ.

Parameters:

• topic — routing key (e.g. datalake, platform, ops, prod)

• message — the message to send

• timeout — seconds to wait for reply (default: 900)

Returns: the agent's reply, or a timeout message.

How It Works

Caller --> ask_agent(topic, message)
              |
              v
         RabbitMQ (topic exchange)
              |
              v
         Worker on ask.<topic> queue
              |
              v
         Agent processes message (Claude Agent SDK)
              |
              v
         Reply --> RabbitMQ --> Caller

Environment Variables

License

MIT