Esco (AgentTalk MCP Server & LAN P2P Bridge)

Esco is a robust, lightweight, and high-performance Model Context Protocol (MCP) integration for the AgentTalk communication bus. It enables multiple AI coding agents (such as Claude, Codex, Clew, OpenCode, OpenClaw, and Hermes-Agent) to communicate and coordinate asynchronously.

By extending AgentTalk's file-based message bus, Esco introduces Peer-to-Peer (P2P) LAN discovery and messaging, allowing agents running on separate machines in the same local network to automatically find and talk to each other without central servers.

---

🚀 Key Features

• P2P LAN Discovery: Automatic peer discovery on local networks using UDP Broadcast on port 9999.
• Direct HTTP Messaging: Messages are routed directly to target peers using dedicated HTTP POST servers bound on ports 18000-18100.
• Identity Registration System: Agents can claim their roster name and role via register_identity() once, mapping process isolation to logical names.
• Blocking Wait Loop: The wait_for_message() tool allows agents to enter a low-overhead, perpetual listening state to wait for incoming commands.
• Decentralized & Serverless: Works fully offline and locally without cloud subscriptions.

---

📁 Repository Structure

Esco/
├── src/agenttalk/             # Core AgentTalk library (message store, CLI, wrapper)
│   └── store.py               # File-backed message bus implementation
├── agenttalk_mcp/             # Our MCP server implementation
│   ├── server.py              # FastMCP server entry point (stdio or sse transport)
│   ├── lan_p2p.py             # UDP broadcast discovery & HTTP P2P messaging engine
│   ├── example_workflow.py    # Working 2-agent task handoff example
│   └── client_simulator.py    # Console-based inter-agent simulation runner
├── AGENT_MCP_GUIDE.md         # Setup manual for Claude, Codex, Clew, etc.
├── .gitignore                 # Excludes __pycache__, runtime .agenttalk/ store, graphify-out/
└── README.md                  # This file

---

🛠️ Getting Started

1. Requirements

Ensure you have Python 3.10+ and the official mcp library installed:

pip install mcp

2. Local Simulation

You can test the MCP server functionality and P2P communication loops locally:

python agenttalk_mcp/client_simulator.py

Or run a real end-to-end multi-agent task handoff (a lead agent assigns work, a worker performs it and reports back, the lead verifies the result):

python agenttalk_mcp/example_workflow.py

The server supports two transports (see AGENT_MCP_GUIDE.md for full details):

• stdio (default) — each agent spawns its own server subprocess.
• sse — one shared server (AGENTTALK_TRANSPORT=sse) that multiple agents/machines connect to over HTTP.

3. Registering the MCP Server in Agent Clients

To register the server for use, configure the command python D:/Projects/Github/Esco/agenttalk_mcp/server.py in your agent configuration.

For Claude Desktop (%APPDATA%\Claude\claude_desktop_config.json):

{
  "mcpServers": {
    "agenttalk-mcp": {
      "command": "python",
      "args": [
        "D:/Projects/Github/Esco/agenttalk_mcp/server.py"
      ]
    }
  }
}

For Codex (~/.codex/config.toml):

[mcp_servers.agenttalk]
command = "python"
args = ["D:/Projects/Github/Esco/agenttalk_mcp/server.py"]

See [AGENT_MCP_GUIDE.md](file:///D:/Projects/Github/Esco/AGENT_MCP_GUIDE.md) for full configuration details.

---

🤝 How Agents Communicate Freely

When configured with this MCP server, agents should adhere to the following workflow:

1. Register Identity: At startup, call register_identity(name="agent_name").
2. Send Message: To communicate, invoke send_message(recipient="target_agent", body="message content").
3. Enter Listen Loop: To wait for incoming responses, block on wait_for_message(). The tool's system instruction enforces that the agent must call this tool at the end of its turn to remain online.