pgEdge Postgres MCP

Install

Works in

• Claude Code
• Claude Desktop
• Cursor
• Windsurf
• GitHub Copilot
• VS Code
• SourceCraft
• Koda
• Coddy

Manual install

sudo apt-get install -y pgedge-postgres-mcp

What it does

pgEdge's Postgres MCP is a Go server that gives an agent SQL queries against PostgreSQL 14 and newer databases. Connection settings live in a YAML file or in the `PGHOST`, `PGPORT`, `PGDATABASE`, `PGUSER` and `PGPASSWORD` variables. The file can describe several databases, and writes are switched on for each one separately with the `allow_writes` flag. By default every query runs in a read-only transaction. Built-in tools: `query_database` runs SQL, `get_schema_info` shows the schema, `count_rows` counts table rows with a filter, `execute_explain` builds a query plan with real timing, `similarity_search` searches by meaning and by words (BM25 and MMR), `generate_embedding` gets a text vector from OpenAI, Voyage AI, Google Gemini or Ollama, and `search_knowledgebase` searches a connected knowledge base. Each tool can be turned off in the config, and a disabled one does not reach the model's list. There is a `system_info` resource and four ready workflows: exploring a database, setting up semantic search, diagnosing a slow query and designing a schema. The server runs over stdio for a local client or over HTTP with users, expiring tokens and OAuth sign-in. The repository also ships a Go chat client that starts the server as a child process and a React web interface for asking the database questions in plain language.

Who it is for. For developers, analysts and database administrators who work with PostgreSQL from Claude Code, Cursor, Windsurf or VS Code with Copilot and want schema, queries and EXPLAIN plans without hand-written SQL. It suits internal teams with trusted users.

Example request

Show the schema of the orders database, count orders with the pending status and analyze with an EXPLAIN plan the query that looks them up by customer email

MCP

Transport

stdio, http

Authentication

not required

Limitations

• The author warns that the server gives the model read access to the whole schema and data of the database and does not recommend it for public-facing applications.
• The built-in chat client and web interface need an Anthropic or OpenAI key or a local Ollama. For the web interface the `llm` section is enabled in the config.
• Embeddings and knowledge base search are off by default. They need a provider: an OpenAI, Voyage AI or Gemini key, or a running Ollama with an embedding model.
• In HTTP mode sign-in is required: a user is created with the `-add-user` flag before the server starts. Installation through Docker Compose works only in HTTP mode and needs `PGEDGE_HTTP_ENABLED=true` in the `.env` file.
• Stdio clients need a binary on disk or the `docker run -i --rm` command.
• Packages in the apt and dnf repositories are built for Debian, Ubuntu, RHEL, Rocky and Alma; on other systems the server is built from source with Go 1.21 or newer.
• The license is the PostgreSQL License.

How to disable. Remove the pgedge entry from the project's `.mcp.json` or from the client settings and restart the agent. For an HTTP server revoke tokens and users: the server rereads their files without a restart.

README in short

The documentation in the docs folder is split into sections: a quick start for the CLI, web interface, Claude Code, Claude Desktop, Cursor, Windsurf and VS Code, deployment in Docker and from source, and a reference for tools, resources and workflows. It separately covers authentication (users, tokens, OAuth), a security checklist, row-level and column-level security, distributed deployment and custom tool definitions. The repository has server, client and web tests, an OpenAPI schema, a changelog and data for a Cursor plugin.