LightRAG MCP tool — README

Summary

• Tool name: query_knowledge_base (registered as query_knowledge_base in lightrag_mcp.py)
  • If you prefer the tool be named lightrag_query, change the decorator to @mcp.tool("lightrag_query") in lightrag_mcp.py.
• Description: Query the internal LightRAG knowledge base for technical documentation. Returns raw text or HTML from the LightRAG server.
• Input schema: QueryInput { query: str, mode: str = 'hybrid' }
• Output schema: QueryOutput { result: str }

---

Quick start (local)

1. Create and activate a virtualenv, then install dependencies:

python -m pip install -r requirements.txt

(Or install packages individually: pip install modelcontextprotocol httpx pydantic python-dotenv)

2. Configure environment variables:

• Copy the example env file and edit as needed:

cp .env.example .env

• Set a system prompt file (optional):

export LIGHTRAG_SYSTEM_PROMPT_FILE="$(pwd)/prompts/light_rag_system.txt"

3. Run the MCP server:

python lightrag_mcp.py

Notes:

• Default FastMCP port is 8000. To change the port, set it in code before mcp.run():

mcp.settings.port = 9680
mcp.run()

• FastMCP exposes streamable HTTP at /mcp (default) and may expose SSE endpoints depending on transport.

---

Example: call the tool from Python (local test)

from importlib import util
import asyncio

spec = util.spec_from_file_location('mod','./lightrag_mcp.py')
mod = util.module_from_spec(spec)
spec.loader.exec_module(mod)

async def demo():
    # Wrap the payload under 'input' when calling via mcp.call_tool
    resp = await mod.mcp.call_tool('query_knowledge_base', {'input': {'query': 'How does the TMS module work?', 'mode': 'hybrid'}})
    print(resp)

asyncio.run(demo())

The call returns a QueryOutput-like structure; depending on transport the response may be a JSON string or an object containing response/result fields.

---

Test connectivity to LightRAG server (direct)

If you need to verify the upstream LightRAG service the MCP calls, you can test it directly (example):

curl -X POST "http://localhost:9621/query" \
  -H "Content-Type: application/json" \
  -d '{"query":"Who owns the TMS module?","mode":"hybrid"}'

This helps verify the configured LIGHTRAG_URL is reachable and responding.

---

Configuration and environment variables

The server looks for these values in order:

1. LIGHTRAG_SYSTEM_PROMPT — inline prompt (env var)
2. LIGHTRAG_SYSTEM_PROMPT_FILE — path to a file containing the prompt
3. prompts/light_rag_system.txt — project default

Recommended .env entries (see .env.example):

LIGHTRAG_SYSTEM_PROMPT_FILE=./prompts/light_rag_system.txt
LIGHTRAG_URL=http://localhost:9621

Security note: Avoid committing sensitive prompts to the repository. Use a secrets manager for production.

---

Transport security and allowed hosts example

If the client (e.g., remote Claude) will connect to your MCP server, restrict allowed hosts in FastMCP settings. Example (add in lightrag_mcp.py before mcp.run()):

# Restrict which remote hosts may call the MCP server (example)
mcp.settings.transport_security.allowed_hosts = ["agents.example.com", "claude.ai"]

Adjust based on your deployment and transport configuration.

---

Registering the tool with clients

Claude Code (VSCode extension / local development)

1. Ensure the MCP server is running locally (e.g., http://localhost:8000).
2. In the Claude Code extension (or your local client), add or point to the server origin and streamable path (usually /mcp).
3. The extension should detect registered tools; verify query_knowledge_base appears in mcp.list_tools().

Remote Claude clients (claude.ai)

1. Expose your local server using a secure tunnel (for testing) or deploy it to a publicly reachable HTTPS endpoint.
   • Example (temporary): ngrok http 8000 → use the provided HTTPS URL
2. Configure transport settings and allowed hosts on the MCP server (see transport security example).
3. Provide the server URL to claude.ai or the remote client in its external tools configuration and confirm the transport type (streamable HTTP vs SSE).

---

Troubleshooting

• Tool not visible in client:

  • Confirm the MCP server is running and reachable.
  • Verify query_knowledge_base appears in mcp.list_tools().
  • Ensure client and server use the same transport (streamable HTTP or SSE).

• Validation errors:

  • Client must send payload matching Pydantic input schema. When calling via the SDK or mcp.call_tool, wrap the payload under input.

• Port already in use:

  • Set mcp.settings.port to a free port before calling mcp.run().

---

Development notes

• lightrag_mcp.py registers the tool as query_knowledge_base. To rename the tool to lightrag_query, either:

  • Change the decorator to @mcp.tool("lightrag_query"), or
  • Update README to reference query_knowledge_base (current state).

• A requirements.txt file is included with the minimal runtime dependencies. Pin versions if you need reproducible installs.

---

Contributing

PRs are welcome. Please avoid committing secrets or sensitive prompts. For changes that alter tool names or schemas, update README and tests accordingly.

---

License

MIT (or replace with your preferred license)