Pgmcp

Visit the GitHub Releases page at github.com/subnetmarco/pgmcp/releases and download the tarball for your OS (Linux, macOS, or Windows). Extract it and place the pgmcp-server binary somewhere on your PATH.

tar xzf pgmcp_*.tar.gz
cd pgmcp_*/
# binaries: pgmcp-server (the MCP server) and pgmcp-client (CLI test tool)

Set DATABASE_URL to your PostgreSQL connection string. Optionally set OPENAI_API_KEY for AI-powered SQL generation — without it the server uses schema introspection only. Other optional variables control the server address, MCP path, and bearer token auth.

export DATABASE_URL="postgres://user:password@localhost:5432/mydb"
export OPENAI_API_KEY="your-openai-api-key"   # optional
export OPENAI_MODEL="gpt-4o-mini"              # optional, default gpt-4o-mini
export HTTP_ADDR=":8080"                       # optional, default :8080
export AUTH_BEARER="your-token"               # optional, enables auth

Run the server binary. It listens on :8080 by default and exposes an MCP endpoint at /mcp. The server connects to your PostgreSQL database and caches the schema for efficient SQL generation.

./pgmcp-server

Use the bundled pgmcp-client to verify the server is working before connecting an AI client. Ask natural language questions and inspect the output in table, JSON, or CSV format.

./pgmcp-client -ask "What tables do I have?" -format table
./pgmcp-client -ask "Who is the customer with the most orders?" -format table
./pgmcp-client -search "john" -format table

Configure Claude Desktop or Cursor to connect to the running PGMCP server via its MCP HTTP endpoint. Alternatively, run the server as a stdio process directly from the client config.

{
  "mcpServers": {
    "pgmcp": {
      "command": "./pgmcp-server",
      "env": {
        "DATABASE_URL": "postgres://user:password@localhost:5432/mydb",
        "OPENAI_API_KEY": "your-openai-api-key"
      }
    }
  }
}