MCP Edge Router Demo

中文说明

Minimal demo of an Edge MCP router (Cloudflare Worker) with backend MCP nodes (FastMCP). It supports progressive disclosure: the gateway only exposes node discovery and tool listing, then forwards tool calls to the selected node.

Project Layout

• edge-worker/ Cloudflare Worker MCP gateway (JSON-RPC over Streamable HTTP)
• nodes/ NodeA–D FastMCP servers (HTTP /mcp)
• mcp/edge_gateway.py Python gateway (local reference)
• test.py Simple LLM agent (OpenAI SDK + MCP)
• start_all.ps1 Start NodeA–D + local Worker

Prerequisites

• Python 3.10+
• Node.js 18+
• Core Python deps in requirements.txt
• Wrangler (via npx is fine)

Install Python Dependencies

conda create -n llm-agent-env python=3.11
conda activate llm-agent-env
pip install -r requirements.txt

Local Run (Windows PowerShell)

1) Start NodeA–D + Worker

powershell -ExecutionPolicy Bypass -File .\start_all.ps1

Note: start_all.ps1 uses conda run -n llm-agent-env .... If you don't use conda, start each node manually.

2) Validate MCP Gateway (JSON-RPC)

# list tools (gateway tools)
curl -s -X POST http://localhost:8787/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":"1","method":"tools/list","params":{}}'

# list nodes (progressive disclosure)
curl -s -X POST http://localhost:8787/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":"2","method":"tools/call","params":{"name":"list_nodes","arguments":{}}}'

# list node tools
curl -s -X POST http://localhost:8787/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":"3","method":"tools/call","params":{"name":"list_node_tools","arguments":{"node_id":"localhost-8001-mcp"}}}'

# call node tool
curl -s -X POST http://localhost:8787/mcp \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":"4","method":"tools/call","params":{"name":"call_node_tool","arguments":{"node_id":"localhost-8001-mcp","tool_name":"math_add","arguments":{"a":2,"b":3}}}}'

Agent Test (OpenAI SDK + OpenRouter-compatible)

1. Create .env at repo root (see .env.example):

OPENROUTER_API_KEY=your_key
OPENROUTER_MODEL=gemini-3-flash-preview
OPENROUTER_BASE_URL=https://api.chataiapi.com/v1
MCP_URL=http://localhost:8787/mcp

2. Run:

python .\test.py

For extensible MCP server config, see mcp/mcp_config.example.json (includes a remote MCP example).

Cloudflare Worker

From edge-worker/:

npx wrangler dev --local

Configure MCP node URLs in edge-worker/wrangler.toml:

[vars]
MCP_NODES = '["http://localhost:8001/mcp","http://localhost:8002/mcp","http://localhost:8003/mcp","http://localhost:8004/mcp"]'

Notes

• Gateway expects MCP nodes to accept application/json, text/event-stream.
• Node descriptions come from FastMCP instructions (see each node).
• Node IDs are derived from URL (stable across restarts).