🐘 Postgres MCP Server

An AI Database Assistant, built on the Model Context Protocol

Connects an AI client to a real PostgreSQL database — schema exploration, safe querying, and LLM-powered reasoning, all through MCP tools.

Docker Hub · Architecture · Installation · Example Tool Calls

What is this?

A production-grade Model Context Protocol (MCP) server that connects an AI client (like Claude) to a real PostgreSQL database, combining direct schema/data access with LLM-powered reasoning (via Groq) for SQL generation, query explanation, optimization, and business insight generation.

Built on a fictional company database ("Orbitals Inc.") — a software consultancy with employees, departments, clients, projects, orders, invoices, support tickets, and meetings — to exercise realistic relational queries and multi-hop relationships.

What makes this an "AI Database Assistant," not just a SQL executor

The database is the source of truth for all facts; the LLM is used for reasoning, translation, and interpretation — never as a substitute for real query execution.

Architecture

MCP Layer (tools/)        → thin, declares tools only
Service Layer (services/) → business logic, orchestration
Repository Layer (repositories/) → raw SQL / SQLAlchemy queries
Database Layer (db/)      → async engine, connection pooling
LLM Client Layer (llm/)   → Groq abstraction, provider-agnostic

Each layer only knows about the one below it — MCP tools never touch the database directly, and the LLM provider is swappable behind an interface (llm/base.py) without touching services.

Folder Structure

postgres-mcp-server/
├── src/postgres_mcp/
│   ├── config.py, server.py, logging_config.py, exceptions.py
│   ├── db/               # async engine, session management
│   ├── models/            # SQLAlchemy models (11 tables)
│   ├── repositories/       # raw SQL against Postgres
│   ├── services/           # business logic
│   ├── llm/                # Groq provider (swappable interface)
│   ├── tools/               # MCP tool registration
│   └── utils/                # SQL injection defenses (sql_guard.py)
├── scripts/               # check_connection, check_llm, seed_database
├── alembic/                # schema migrations
├── tests/
│   ├── unit/                # sql_guard, schema_service, insight_service
│   └── integration/          # real Neon queries
├── Dockerfile, docker-compose.yml
└── alembic.ini

Database Schema

11 tables modeling a software consultancy: departments, employees (self-referencing manager hierarchy), clients, projects, project_assignments (many-to-many), products, orders, invoices, support_tickets, meetings, meeting_attendees (many-to-many). Full relational integrity via foreign keys, check constraints (e.g., positive order quantities, valid status enums), and indexes on frequently-filtered columns.

Installation

Prerequisites

• Python 3.12+

• A Neon PostgreSQL database (free tier works)

• A Groq API key (free tier works)

• Docker (optional, for containerized runs)

Setup

git clone <your-repo-url>
cd postgres-mcp-server
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env
# fill in .env with your real Neon connection string and Groq API key

Environment Variables

Neon Setup

1. Create a free account and project at neon.tech

2. Copy the connection string from your project dashboard

3. Convert it to async form: change postgresql:// to postgresql+asyncpg://, and sslmode=require to ssl=require, dropping any channel_binding param

Database Setup

alembic upgrade head              # creates all 11 tables
python scripts/seed_database.py   # populates realistic fictional data

Running Locally

python -m postgres_mcp.server

This starts the MCP server over stdio transport. To test it interactively:

npx @modelcontextprotocol/inspector@latest python -m postgres_mcp.server

Public MCP Endpoint

The hosted MCP server is available at:

https://aman-postgres-mcp-server.fastmcp.app/mcp

Running with Docker

docker build -t postgres-mcp-server:latest .
docker run -i --env-file .env postgres-mcp-server:latest

Or via Docker Compose:

docker compose up --build

Pre-built image available on Docker Hub:

docker pull amankhan1009/postgres-mcp-server:latest

Testing

pytest -m "not integration" -v   # unit + mocked tests (fast, no DB needed)
pytest -m integration -v          # integration tests (hits real Neon)
pytest -v                          # everything

22 tests passing: 20 unit/mocked (including full sql_guard injection-defense coverage) + 2 integration tests against real seeded data.

Example Tool Calls

list_tables()

["clients", "departments", "employees", "invoices", "meeting_attendees",
 "meetings", "orders", "products", "project_assignments", "projects", "support_tickets"]

business_insights("which department has the highest average employee salary?")

The department with the highest average employee salary is Engineering, with an average salary of $101,125.

find_related_tables("employees", max_depth=2)

{
  "table_name": "employees",
  "related_tables": {
    "departments": 1, "project_assignments": 1, "meeting_attendees": 1, "support_tickets": 1,
    "projects": 2, "clients": 2, "meetings": 2
  }
}

Security

• All user/LLM-supplied SQL passes through sql_guard.py: single-statement enforcement, forbidden keyword/function blocking, and a hard row limit

• Every query runs inside a READ ONLY Postgres transaction as a final safety net

• Table names from tool inputs are validated against a live whitelist (list_tables()) before being used in any query, since identifiers can't be parameterized like values

• Secrets are never baked into the Docker image — passed only at runtime via --env-file

• Container runs as a non-root user

Future Improvements

• Optional, separately-hardened write capability (structured insert_row-style tools with strict table/column whitelisting, audit logging, and confirmation steps) — deliberately out of scope for this version, which is read-only by design

• Support for a second LLM provider (OpenAI/Anthropic) via the existing LLMProvider interface

• Query result caching for repeated describe_table/list_tables calls

• Rate limiting on execute_select and LLM-powered tools

Built as a portfolio project to demonstrate production-grade backend architecture, PostgreSQL fluency, and safe AI-database integration via MCP.