Awesome AI Agents › MCP Server Framework

subnetmarco/pgmcp

⭐ 540 Go repository created 2025-09-16

PGMCP is a Model Context Protocol server that puts an existing PostgreSQL database behind a natural language interface for LLM clients. It connects to a database through a standard connection string, reads and caches the schema, and answers questions posed in plain English by generating SQL against that schema. Any MCP-compatible client can drive it, including Cursor, Claude Desktop and VS Code extensions, and the project also ships its own command line client for direct queries and for scripting. Nothing in the target database has to change: no schema modifications, no extra tables, and mixed-case table and column names are handled correctly. Query generation is delegated to an OpenAI-compatible model, configured through an API key and model name, with gpt-4o-mini as the default; the server degrades gracefully when the model call fails, and a free-text search mode works across all text columns without needing the model at all. Safety is the main design constraint. Access is strictly read-only, writes are blocked, generated statements pass through an SQL guard and input validation, queries are recorded in an audit log, and an optional bearer token protects the HTTP endpoint. Large result sets are paginated and streamed automatically so that memory use stays bounded, and connections come from a pool. Results can be rendered as an ASCII table, JSON or CSV. The server is a single Go binary served over streamable HTTP at a configurable address and path. It suits data teams and developers who want conversational analytics or ad-hoc exploration over production data through an agent, without granting the agent write access or building a bespoke query layer.

https://github.com/subnetmarco/pgmcp

mcpmcp-serverpostgrespostgresqlnatural-language-to-sqldatabaseanalyticsdata-analysisagentic-aigo