MCP Servers

LangGraph + MCP Demo

End-to-end demo that spins up:

• FastAPI service that also bootstraps an MCP server.
• MCP tools backed by Postgres (get_user_balance, get_last_logins).
• LangGraph agent that calls those MCP tools via the Model Context Protocol.
• Docker Compose orchestration for the API + Postgres.

Prerequisites

• Docker + Docker Compose
• (Optional) Python 3.11+ if running locally without containers
• OpenAI API key (or compatible endpoint) for the agent LLM

Quick start (Docker)

cp .env.example .env   # set OPENAI_API_KEY
docker compose up --build

Services:

• API: http://localhost:8000
• MCP server (HTTP transport): http://localhost:9000
• Postgres: localhost:5432 (db: mcp_demo, user: postgres, password: postgres)

API checks

• Health: curl http://localhost:8000/health
• Balance: curl http://localhost:8000/balance/alice
• Logins: curl 'http://localhost:8000/logins/alice?limit=3'
• Agent: curl -X POST http://localhost:8000/agent -H 'Content-Type: application/json' -d '{"question":"What is Alice\'s balance and last 2 logins?"}'

Local dev without Docker

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill OPENAI_API_KEY
python -m app.main     # runs FastAPI+MCP; uvicorn launched inside

Project layout

• app/main.py – FastAPI app, spins MCP server on startup.
• app/mcp_server.py – MCP tool definitions (balance + logins).
• app/agent.py – LangGraph agent that consumes MCP tools.
• app/db.py / app/services.py – async DB access helpers.
• db/init.sql – schema + demo seed.
• docker-compose.yml / Dockerfile – container setup.

Notes

• MCP transport is streaming-http on port 9000 for simplicity.
• The agent uses gpt-4o-mini by default; override via OPENAI_MODEL.
• Database seeding happens automatically on startup.