PostgreSQL Model Context Protocol (PG-MCP) Server

What is PostgreSQL Model Context Protocol (PG-MCP) Server?

It is an MCP server for PostgreSQL databases that provides a comprehensive API for AI agents to discover, connect to, query, and understand PostgreSQL databases through MCP's resource-oriented architecture. It is built for AI agents and developers who need to integrate natural language interfaces with PostgreSQL.

How to use PostgreSQL Model Context Protocol (PG-MCP) Server?

Install via Docker or manual setup with Python 3.13+. Run python -m server.app after installing dependencies with uv sync --frozen. Clients connect to the SSE endpoint (default http://localhost:8000/sse) and invoke tools like connect, pg_query, and pg_explain using a secure connection ID.

Key features of PostgreSQL Model Context Protocol (PG-MCP) Server

• Multi-database support with simultaneous connections.
• Rich catalog information from table/column descriptions.
• Query execution plan analysis via pg_explain tool.
• Built-in extension context for PostGIS and pgvector.
• Read-only mode enforced via transaction settings.
• Connection pooling and secure opaque connection IDs.

Use cases of PostgreSQL Model Context Protocol (PG-MCP) Server

• AI agents exploring database schema and sample data.
• Natural language to SQL conversion using Claude CLI.
• Automated database introspection for documentation or analysis.
• Query performance analysis with execution plan tool.

FAQ from PostgreSQL Model Context Protocol (PG-MCP) Server

What are the runtime requirements?

Python 3.13+ and a PostgreSQL database. The server can be run via Docker or manually.

How do I connect a database?

Use the connect tool with a PostgreSQL connection string to obtain an opaque connection ID for subsequent queries.

Is the server read-only?

Yes, the server runs in read-only mode enforced via transaction settings.

What PostgreSQL extensions are supported?

Built-in context for PostGIS and pgvector is provided, and additional extensions can be added via YAML files.

How is security handled?

Connection details are never exposed in resource URLs; only opaque connection IDs are used. Credentials are sent only once during the initial connection.