GCP Postgres MCP Tools

FastAPI app that connects to your GCP PostgreSQL database and exposes query tools over REST and MCP.

Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env

Put your Cloud SQL URL in .env. If you already use a local Cloud SQL Auth Proxy, point at that host and port:

DATABASE_URL=postgresql+asyncpg://USERNAME:PASSWORD@127.0.0.1:5439/DATABASE_NAME
USERS_TABLE=tbl_users
USERS_ORDER_COLUMN=date_joined

This app is read-only against Cloud SQL. Every production session sets default_transaction_read_only and SET TRANSACTION READ ONLY. REST tools are GET-only. POST is allowed only on /mcp (protocol) and /agent/chat (the agent loop). Agent traces are written to a separate local SQLite file, never to evolve_production.

get_last_10_users uses tbl_users. get_last_5_user_activities uses tbl_user_activities. Password columns are never returned.

For production, also use a Cloud SQL user that has SELECT only.

Related MCP server: MCP PostgreSQL Server

Run

uvicorn app.main:app --reload --port 8000

• Health: GET /health

• Docs: http://127.0.0.1:8000/docs

• MCP: http://127.0.0.1:8000/mcp

• Agent: POST /agent/chat

• Traces: GET /agent/traces and GET /agent/traces/{trace_id}

Agentic loop and traces

The agent is a ReAct loop: think → call a read-only tool → observe → repeat → final answer.

curl -s http://127.0.0.1:8000/agent/chat \
  -H 'Content-Type: application/json' \
  -d '{"message":"Get the last 10 users"}'

Each run stores a structured trace (user_input, LLM tool choices, tool results, final answer) in data/traces.db.

The agent uses the Gemini Developer API (Google AI Studio key), not Vertex AI and not OpenAI.

GEMINI_API_KEY=AIza...
GEMINI_MODEL=gemini-3.6-flash

Get a key at aistudio.google.com/apikey. GOOGLE_API_KEY also works.

These are not the same as a Vertex AI key. Gemini keys start with AIza and do not need a GCP project or region.

Optional: LangSmith via LANGSMITH_TRACING=true and LANGSMITH_API_KEY.

Without GEMINI_API_KEY, /agent/chat returns 503. REST and MCP read tools still work.

Each chat request is capped so the agent cannot loop or retry forever:

Tools

Password and hash columns are never returned.

Cursor MCP config

{
  "mcpServers": {
    "gcp-postgres-tools": {
      "url": "http://127.0.0.1:8000/mcp"
    }
  }
}