pgembedReal PostgreSQL, any version, no Docker

Run named Postgres instances side by side, query them with readable results, inspect their schema, and hand them to your AI assistant over MCP. One .NET tool.

Get Started

MCP Server

View on GitHub

Any version, side by side

Postgres 9.6 to 18 from the official binaries, on Windows, macOS and Linux (glibc and musl, x64 and arm64). Several versions at once, each a named instance that keeps its port.

Instances

Minimal, small downloads

Just the server, initdb and pg_ctl, nothing else. About 14 MB on Linux and 24 MB on Windows, against 112 to 154 MB for the official Docker image. Downloaded once into a shared cache, checksum-verified.

What you get

No Docker, no install

Works on laptops and CI runners that cannot run Docker or install packages. Install with npm or as a .NET tool; nothing else to set up.

Why pgembed

SQL you can read

Tables that fit the terminal, psql-style text for pipes, JSON with nested json values, CSV and Markdown. Errors point at the position, with the server's hint.

Run SQL

Safe by default

Statements that change the schema or existing rows are listed and confirmed before anything runs, from the CLI and from MCP. UPDATE or DELETE without WHERE is called out.

How it decides

A throwaway Postgres for AI assistants

pgembed mcp gives Claude, Copilot and other MCP clients 18 tools to start instances, run and explain SQL, and read schemas, with read-only mode and confirmation.

Set it up

Shared with your tests

Instances live where MysticMind.PostgresEmbed looks, so a test attaches to `pgembed start app` in milliseconds, or starts it on its own.

From .NET tests

Fast

A native executable per platform; a query against an instance takes about 25 ms end to end, faster than psql. Results stream in under 40 MB, for one row or five million.

The numbers

The real PostgreSQL parser

Scripts split correctly around dollar quotes, syntax checks and formatting work offline, and each statement is summarized before it runs.

Format and check