Best MCP Servers for Claude Desktop (2026)
Last updated 2026-10-09
Quick answer
The best MCP servers for Claude Desktop in 2026 are Filesystem, GitHub, Playwright, Context7, Memory and Sequential Thinking — depending on what you want Claude to do. Claude Desktop talks to local servers over stdio through claude_desktop_config.json, so the servers below install with one config block and a restart. Start with two or three, not ten.
Top 10 MCP servers for Claude Desktop
Ranked by practical usefulness, not GitHub stars. Free means the MCP server itself is free; connected services may have separate costs.
- ▸Filesystem — read, write and organise local files (free)
- ▸GitHub MCP — repositories, issues, pull requests (free server)
- ▸Playwright — browser automation and UI testing (free)
- ▸Context7 — current, version-specific programming documentation (free tier)
- ▸Memory — persistent structured knowledge across sessions (free)
- ▸Sequential Thinking — structured multi-step reasoning (free)
- ▸Chrome DevTools — browser debugging and performance (free)
- ▸Fetch — retrieve web page content (free)
- ▸Git — local repository history and changes (free)
- ▸Brave Search — live web search (API limits/pricing)
How we picked
Claude Desktop was the first MCP client and remains the reference implementation, so the servers that work best there are the ones published by the MCP project itself and platform vendors with clean install steps. Five of the ten above — Filesystem, Memory, Git, Fetch and Sequential Thinking — are official reference implementations maintained in the MCP project's own repository.
For developers, prioritise servers that let Claude actually perform tasks rather than simply retrieve information: file access, repository operations and a real browser close the loop between suggestion and action.
A sensible starter stack
You do not need all ten on day one. Each server adds tools, and every tool spends context window. A practical first three:
- ▸Filesystem, scoped to one project directory — lets Claude read and edit real files
- ▸GitHub — issues, pull requests and code search without leaving the chat
- ▸Brave Search or Fetch — one server for fresh, citable web facts
How to install them
Every server on this list connects through Claude Desktop's config file: Settings → Developer → Edit Config, then paste the server's mcpServers JSON block and restart the app. Each server page in this directory shows a copy-paste config block — for example the Filesystem server page has the exact JSON with safe directory scoping.
Full walkthrough: Claude Desktop MCP setup guide.
Safety notes
Filesystem and Git servers can change your machine. Scope Filesystem to the single project directory you are working in — never your home folder. Start GitHub with read-only scopes and keep a human in the loop for merges and pushes. Browser servers (Playwright, Chrome DevTools) can see whatever the browser can, including logged-in sessions; use a dedicated browser profile. The MCP security guide covers the full checklist.
The picks, one by one
Official reference server for scoped local file access — read, write and organise files inside directories you allow.
Best for: Letting Claude work on real project files instead of pasting code into chat.
Watch out: Scope it to one project directory, never your home folder.
Official GitHub server: repositories, issues, pull requests and code search from the chat window.
Best for: Developers triaging issues and reviewing PRs without switching tabs.
Watch out: Can write to repositories — start with read-only scopes.
Microsoft's Playwright server drives a real browser: navigation, forms, screenshots and assertions.
Best for: UI testing and end-to-end checks from inside Claude Desktop.
Watch out: It sees logged-in sessions — run it in a dedicated browser profile.
Upstash's Context7 injects current, version-specific library docs into context, killing hallucinated APIs.
Best for: Any coding work with libraries that move faster than model training data.
Watch out: Adds tokens per lookup — fine for daily use, skip it for trivial chats.
Reference knowledge-graph server that persists facts across conversations.
Best for: Long-running projects where context must survive restarts.
Watch out: Only as accurate as what was stored — prune it.
Reference server for structured, step-by-step planning on hard problems.
Best for: Multi-step tasks where jumping straight to an answer fails.
Watch out: Extra reasoning rounds mean slower, longer answers.
Google's official Chrome DevTools server attaches to a live tab — console errors, network, screenshots.
Best for: Debugging web apps with real telemetry instead of pasted stack traces.
Watch out: Same session-visibility caveat as Playwright.
Reference server that retrieves a URL and converts it to clean markdown for the model.
Best for: Reading a specific page or doc on demand.
Watch out: No search — pair it with Brave Search for discovery.
Official Git server for local repository operations: status, diffs, log and commits.
Best for: Inspecting and committing to local repos with Claude's help.
Watch out: Push and history rewriting deserve a human review.
Live web search tuned for agents, returning fresh citable results.
Best for: Facts the model cannot know: news, prices, recent releases.
Watch out: Requires a Brave API key with monthly free tier limits.
Frequently asked questions
How many MCP servers should I run in Claude Desktop at once?＋
Start with two or three. Every server adds tools to Claude's context, which costs window space and can confuse tool selection. Add more only when a real task needs them.
Do these servers cost money?＋
The servers themselves are free and open source. Some connect to paid APIs — Brave Search has a free tier with monthly limits, and GitHub is free against GitHub's own rate limits.
Where does Claude Desktop's config file live?＋
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json. Windows: %APPDATA%\Claude\claude_desktop_config.json. Open it from the app via Settings → Developer → Edit Config.
Which servers are official reference implementations?＋
Filesystem, Memory, Git, Fetch and Sequential Thinking are maintained in the MCP project's own repository (modelcontextprotocol/servers). GitHub, Playwright, Chrome DevTools and Context7 are published by their platform vendors.