Ask Your Database — MCP Server

A local MCP server that lets Claude Desktop or Claude Code answer plain-English questions against a Postgres database — e.g. "which customers spent less than $2,000 in the last year?" — by calling real tools, not by guessing.

Tested end-to-end: schema discovery, multi-step joins, and a blocked write all verified against a live Postgres instance.

What's here

mcp-demo/
├── server.py           # MCP server — exposes list_tables, describe_table, query
├── db.py                # connection + safety layer (read-only role, timeout, write-block)
├── seed.sql              # schema + sample data + read-only role setup
├── docker-compose.yml    # spins up Postgres with the seed data pre-loaded
├── requirements.txt       # mcp[cli], psycopg2-binary
└── README.md

Related MCP server: Postgres MCP Server

1. Start Postgres

cd mcp-demo
docker compose up -d

This starts Postgres on localhost:5432, creates the demo database, loads the sample schema (customers, subscriptions, support_tickets), and creates a mcp_readonly role that can only SELECT.

No Docker? Run psql against any local Postgres instance and load seed.sql directly: psql -U postgres -d demo -f seed.sql (creates the database first if needed).

2. Set up Python

python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

3. Point the server at your database

The server reads connection details from environment variables (defaults shown match seed.sql and docker-compose.yml, so you likely don't need to change anything for local testing):

4. Smoke-test it standalone (optional but recommended)

export PGHOST=localhost PGPORT=5432 PGDATABASE=demo PGUSER=mcp_readonly PGPASSWORD=readonlypass
python3 server.py

It should sit there quietly waiting on stdio — that's correct, it's not meant to print anything until a client (like Claude Desktop) talks to it. Ctrl+C to stop.

5. Connect it to Claude Desktop

Open Claude Desktop's config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add this server (use the absolute path to server.py on your machine):

{
  "mcpServers": {
    "ask-your-database": {
      "command": "/absolute/path/to/mcp-demo/venv/bin/python3",
      "args": ["/absolute/path/to/mcp-demo/server.py"],
      "env": {
        "PGHOST": "localhost",
        "PGPORT": "5432",
        "PGDATABASE": "demo",
        "PGUSER": "mcp_readonly",
        "PGPASSWORD": "readonlypass"
      }
    }
  }
}

Restart Claude Desktop. You should see "ask-your-database" listed under the 🔌 tools/connectors icon in a new chat.

6. Connect it to Claude Code

From the project directory:

claude mcp add ask-your-database \
  --env PGHOST=localhost --env PGPORT=5432 --env PGDATABASE=demo \
  --env PGUSER=mcp_readonly --env PGPASSWORD=readonlypass \
  -- /absolute/path/to/mcp-demo/venv/bin/python3 /absolute/path/to/mcp-demo/server.py

7. Try it

Ask Claude, in plain English:

"Which customers churned last month, and what plan were they on?"

"Which customers spent less than $2,000 total in the last year?"

Claude will call list_tables and describe_table on its own to learn the schema (nothing about it is hardcoded in the prompt), then write and run the SQL itself via the query tool.

Try a write, too — it's blocked on purpose:

"Delete all customers with status 'churned'."

Adapting this to your real database

Swap the seed data for your actual schema, and point PGHOST/PGDATABASE at your real Postgres instance — using a read-only role, same as here. Never point this server at a production database with a role that can write. db.py's query logic doesn't need to change at all.

Troubleshooting

• Server doesn't show up in Claude Desktop — check the path in claude_desktop_config.json is absolute, not relative, and that you restarted the app after editing the config.

• permission denied on every query, even simple ones — check the mcp_readonly role's grants ran (last block of seed.sql); rerun that block if you dropped and recreated the database.

• Connection refused — confirm Postgres is actually running (docker compose ps) and the port matches PGPORT.