Project 3 — Custom MCP Server + LangGraph Agent (Fully-Free Stack)

A personal knowledge assistant built on the Model Context Protocol (MCP). A custom FastMCP server exposes tools (semantic note memory + web search); a LangGraph ReAct agent discovers and calls those tools over HTTP.

User Query -> LangGraph Agent -> MultiServerMCPClient -> MCP Server (FastMCP)
                                                          |-- Qdrant (notes)  + FastEmbed (local)
                                                          |-- Tavily (web)

Free stack (no paid keys)

The note-memory tools (add_note, list_notes, search_notes) need no API key at all — embeddings run locally. Only the agent's LLM needs a (free) OpenRouter key.

Related MCP server: memory-mcp

Status on this machine

1. Add your free OpenRouter key

Get one at https://openrouter.ai/keys, then put it in .env:

OPENROUTER_API_KEY=sk-or-...
# OPENROUTER_MODEL=meta-llama/llama-3.3-70b-instruct:free   # optional override
TAVILY_API_KEY=                                              # optional web search

The agent is a tool-calling ReAct agent, so the OpenRouter model must support function/tool calling. Good free options: meta-llama/llama-3.3-70b-instruct:free, qwen/qwen-2.5-72b-instruct, deepseek/deepseek-chat. If a model ignores tools, switch OPENROUTER_MODEL.

2. Start the MCP server (own terminal)

venv/Scripts/python mcp_server.py

Serves MCP at http://localhost:8001/mcp.

3. Verify without keys (optional)

venv/Scripts/python test_connection.py   # tool discovery + list_notes
venv/Scripts/python test_memory.py       # add -> list -> semantic search

4. Run the agent (needs OpenRouter key)

venv/Scripts/python mcp_agent.py "Save a note titled 'RAG Tips': Always use hybrid search"
venv/Scripts/python mcp_agent.py "What did I learn about retrieval?"
venv/Scripts/python mcp_agent.py "What notes do I have?"
venv/Scripts/python mcp_agent.py "Search the web for news about LangGraph 2026"   # needs Tavily

Compatibility fixes applied vs. the original handout

The handout code targets older library versions. Updated for current releases:

1. Embeddings -> local FastEmbed (mcp_server.py). No OpenAI key; EMBED_DIM changed 1536 -> 384 to match bge-small-en-v1.5.

2. Agent LLM -> OpenRouter via ChatOpenAI(base_url=...) (mcp_agent.py), replacing init_chat_model("anthropic:...").

3. MultiServerMCPClient is not a context manager anymore (langchain-mcp-adapters 0.1.0+) — instantiated directly, then get_tools().

4. qdrant.search() -> qdrant.query_points(...).points (qdrant-client 1.12+).

Inspect the server interactively (optional)

npx @modelcontextprotocol/inspector http://localhost:8001/mcp