agent-mesh-mcp

MCP server for inter-agent communication via RabbitMQ. Agents ask each other questions routed by topic.

Install

npm install -g agent-mesh-mcp

Usage

Claude Code (~/.claude.json)

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

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "agent-mesh": {
      "command": "agent-mesh-mcp",
      "env": {
        "RABBITMQ_URL": "amqp://user:pass@your-rabbitmq:5672/",
        "AGENT_NAME": "cursor"
      }
    }
  }
}

npx (no install)

{
  "mcpServers": {
    "agent-mesh": {
      "command": "npx",
      "args": ["-y", "agent-mesh-mcp"],
      "env": {
        "RABBITMQ_URL": "amqp://user:pass@your-rabbitmq:5672/"
      }
    }
  }
}

Tool

ask_agent(topic, question)

Ask another agent a question. The question is published to a RabbitMQ topic exchange and routed to the agent listening on that topic.

Parameters:

• topic — routing key (e.g. backend, frontend, ops, qa, devops)
• question — the question to ask

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

How It Works

Caller ──► ask_agent(topic, question)
              │
              ▼
         RabbitMQ (topic exchange)
              │
              ▼
         Worker on ask.<topic> queue
              │
              ▼
         Agent processes question
              │
              ▼
         Reply ──► RabbitMQ ──► Caller

Each agent is a worker process listening on its topic queue. When a question arrives, the worker runs an AI model (e.g. Claude) to answer it, then publishes the reply back.

Environment Variables

License

MIT