Cursor caps how many MCP tools it loads (~40), so the goal is not "install everything" — it is picking four to six servers that earn their slot. These ten are the ones we keep coming back to, grouped by job.
Context & docs
- Context7 — injects version-specific docs for hundreds of frameworks into the prompt, killing hallucinated APIs. The single highest-leverage server for coding in Cursor.
- Filesystem — read and write project files. The backbone of almost every setup. See the filesystem setup guide.
Code & repos
- GitHub — issues, PRs, code search, review. See GitHub MCP server setup.
- Git — local history, diffs and blame without leaving the agent.
Databases
- Postgres — query your schema and data safely. See Postgres MCP setup.
- DBHub — one server for MySQL, SQLite, SQL Server and more; see MySQL MCP setup.
Web & automation
- Brave Search / Exa — live web grounding for current information.
- Playwright — drive a real browser to test and scrape. See Playwright setup.
- Fetch — pull and clean a single URL into context.
Reasoning aids
- Memory — persist facts across sessions so the agent stops forgetting project context.
How to choose (the 40-tool rule)
Each server adds several tools, and Cursor stops loading past its ceiling. Pick by your actual workflow: a web dev wants Filesystem + GitHub + Context7 + Postgres + Playwright; a data engineer swaps in DBHub and BigQuery. Disable what you are not using in Settings → MCP.
If a server will not load, it is usually the cap or a PATH issue — see Cursor MCP not working.
Ready-made stacks
Rather than assemble by hand, grab a tested combination from our curated loadouts, or see the backend developer stack. New to Cursor MCP? Start with the Cursor MCP setup guide.