Best MCP Servers for Cursor (2026)
Last updated 2026-10-09
Quick answer
For Cursor, the best MCP servers are GitHub, Context7, Playwright, Supabase and Filesystem — they give the agent repository access, current documentation, browser testing, databases and local files. Most are free and open source. Cursor supports both one-click installs from its MCP settings and manual .cursor/mcp.json configuration.
Top 10 MCP servers for Cursor
Ranked by development usefulness, not benchmarks. 'Free tier' means the server is free but the connected service has limits.
- ▸GitHub MCP — repository management, issues, pull requests (free / open source)
- ▸Context7 — up-to-date library and framework documentation (free tier)
- ▸Playwright MCP — browser automation, UI testing, screenshots (free / open source)
- ▸Supabase MCP — PostgreSQL databases and backend development (free tier)
- ▸Filesystem MCP — local files and directories (free / open source)
- ▸Figma MCP — design-to-code workflows (plan dependent)
- ▸Sentry MCP — production errors and debugging (free tier)
- ▸Sequential Thinking MCP — structured reasoning and planning (free / open source)
- ▸Exa MCP — web research and technical information (usage dependent)
- ▸Notion MCP — product specs and documentation (plan dependent)
How we picked
Cursor is a coding agent first, so we ranked servers by what they fix in real development loops: stale documentation (Context7), no repo access (GitHub), inability to see the running app (Playwright), missing backend (Supabase) and lost plans (Sequential Thinking). Publishers with official status — Microsoft for Playwright, GitHub for github-mcp-server — beat community forks on safety and update cadence.
For most Cursor development workflows, start with GitHub + Context7 + Playwright and add others only when a task needs them.
The recommended starter stack, one by one
GitHub MCP Server — review changes and work issues and pull requests without leaving the editor.
Context7 — retrieves current, version-specific documentation to reduce outdated-API usage and invented code.
Playwright MCP — lets Cursor operate a real browser to test pages, forms, navigation and responsive behaviour.
Supabase MCP — useful when building full-stack features against a database; prefer a development database over production.
How to install MCP servers in Cursor
Cursor supports one-click installation from its directory and manual JSON configuration. UI route: Cursor Settings → MCP → Add new MCP server, search for the server, authorise and verify. Manual route: create .cursor/mcp.json in your project directory (or ~/.cursor/mcp.json for global) and paste the server's mcpServers block — every server page in this directory includes one.
Full walkthrough with a working example: Cursor MCP setup guide.
Safety notes
Database servers deserve special care in an IDE agent: point Supabase or Postgres servers at a dev database, never at production credentials sitting in your .env. Playwright sees logged-in browser sessions — use a separate profile. And review agent-driven commits before pushing, the same way you would review a junior developer's work. More: MCP security best practices.
The picks, one by one
Official GitHub MCP server: repositories, issues, pull requests and code search inside the editor.
Best for: Daily repo work — reviewing diffs and closing issues without tab-switching.
Watch out: Can write to repos; start with read-only scopes.
Version-specific library documentation injected on demand — the biggest accuracy upgrade for generated code.
Best for: Projects on fast-moving frameworks where training data lags reality.
Watch out: Each lookup costs context tokens.
Microsoft's official Playwright server: Cursor drives a real browser for tests and screenshots.
Best for: Verifying UI changes end-to-end before you open a PR.
Watch out: Runs with your browser profile's access — isolate it.
Community-maintained Supabase server for schema inspection, queries and project management.
Best for: Full-stack features built against a Supabase backend.
Watch out: Use a dev project; production data deserves human hands.
Reference server for scoped local file access — read and edit files in allowed directories.
Best for: Agents that need to touch files outside the open workspace.
Watch out: Scope to the project; not your home directory.
Figma's official server: read designs and design context for design-to-code workflows.
Best for: Turning Figma frames into components with real spacing and tokens.
Watch out: Requires a Figma plan that includes MCP access.
Official Sentry server: pull real production errors and traces into the debugging loop.
Best for: Fixing bugs with actual stack traces and breadcrumbs.
Watch out: Read-only by default — keep it that way.
Reference server for decomposing hard tasks into structured steps.
Best for: Refactors and multi-file changes with many moving parts.
Watch out: Slower answers; skip for small edits.
Neural web search built for agents — semantic queries over keyword matching.
Best for: Technical research: docs, RFCs, comparison posts.
Watch out: Usage-based pricing beyond the free tier.
Official Notion server: search, read and update pages and databases.
Best for: Pulling product specs and writing docs back from the editor.
Watch out: Wide token scopes — audit what it can touch.
Frequently asked questions
Does Cursor support remote MCP servers?＋
Yes. Cursor accepts both local stdio servers (command-based) and remote servers (URL-based, Streamable HTTP). Remote servers skip local installation entirely — paste the URL in Settings → MCP.
Where is Cursor's MCP config file?＋
Per-project: .cursor/mcp.json in the project root. Global: ~/.cursor/mcp.json. The settings UI edits the same file. Every server page on MCPNav shows a ready-to-paste mcpServers block.
What is the difference between Cursor's built-in tools and MCP servers?＋
Built-in tools cover the open workspace: file edits, terminal, codebase search. MCP servers extend the agent beyond that — to GitHub issues, live browsers, databases and external docs.
Do MCP servers slow Cursor down?＋
Enabled servers add tool definitions to the model's context on every request, which costs tokens and can dilute tool selection. Enable servers per project and disable the ones you are not using.