Summary

SSE endpoints, so one stack serves every client on the machine.

Because these listen on ports rather than stdio, the same running container answers Claude, Cursor and anything else at once — no duplicate process per client. Treat it as a starting point rather than a product: it is published as an internal collection, and the Parquet server is explicitly still in progress.

What it is

A collection of MCP servers built for one team's internal use and published as-is. Each exposes a different toolset — browser automation, a vector store, a relational database — through a natural-language interface.

What you get

Puppeteer: control an automated browser through Chrome or Browserless
Postgres vector store: search and interact with a PostgreSQL-backed vector database
MySQL: query and work with a MySQL database in plain language
Parquet, marked in progress: convert between JSON and Parquet

Requirements

Clone the repo and bring the servers up with docker compose, or run one directly in development with `tsx src/[server]` where the module is `puppeteer`, `mysql` or `pgvector`. Clients connect over SSE — the sample config points Puppeteer at `http://localhost:8000/sse`, pgvector at `http://localhost:8001/sse` and MySQL at `http://localhost:8002/sse`. Database connection details are yours to supply.

Setup effort

One command — npx -y mcp-server