Best MCP servers in 2026: the ones worth actually installing
The MCP ecosystem passed 17,000 public servers. You need about eight. Here is which eight, and why the rest are noise.
The Model Context Protocol won. It was donated to the Linux Foundation's Agentic AI Foundation in December 2025, and OpenAI and Google DeepMind both adopted it. Counting servers is harder than it sounds — the official registry curates to roughly 10,000 records while directories like mcp.so and Glama index north of 20,000, because they count different things. Either way the abundance is the problem: every server you install costs context on every turn, so the useful question is not which exist but which few earn their place.
The 2026 default is remote, not local
GitHub, Vercel, Linear, Notion, Supabase, Stripe and Figma all publish OAuth-secured hosted endpoints now, so you skip the local install entirely. Zuplo's State of MCP survey found 59% of builders on Streamable HTTP against 34% still on stdio. If a setup guide tells you to clone a repo and run a Node process, check whether the vendor has since shipped a hosted endpoint.
The ones almost everyone should have
Context7 is the most recommended server in developer communities right now, and deservedly. It pulls live documentation so the model stops hallucinating API signatures that were deprecated two versions ago. If you install exactly one thing from this list, install this.
GitHub gives an agent issues, pull requests and code search — the connective tissue between what you ask for and where the work lives. Filesystem and Git are the local pair that make an agent useful on a real repository rather than a chat transcript.
Sentry closes a loop most setups leave open — the agent that wrote the code can see the error it caused in production. Worth knowing it is a much smaller project than the others here, so treat it as a useful addition rather than a staple.
Browser automation: pick the right one
This is where people install three servers that do overlapping jobs. The distinction is real and worth learning once.
- Playwright MCP — repeatable end-to-end assertions across browsers. Microsoft-maintained, the broadest cross-browser support. Reach for it when you want a test.
- Chrome DevTools MCP — interactive debugging of one live Chrome instance: performance traces, network inspection. Official Google server, and by GitHub stars the most adopted of the four. Use it when you need evidence, not a test.
- Puppeteer MCP — quick local Chromium screenshots and DOM extraction. Now archived upstream; reach for Playwright instead unless you have a reason.
- Browserbase — managed sessions with replay and persistence, for when browser automation goes to production.
Chrome DevTools MCP is for debugging one browser. Playwright MCP is for asserting across many. Installing both because they sound similar just costs you context.
Data and backend
Supabase gives an agent visibility into a Postgres backend — schema, auth, storage, edge functions — which is why it became the default for people building on it. Neon, ClickHouse and MongoDB cover the same job for their own stacks. Cloudflare exposes DNS, Workers, Pages deployments, caching and firewall rules, which is more useful than it sounds the first time an agent diagnoses a caching problem for you.
Search and retrieval
Exa, Tavily and Firecrawl all fetch the web, with different emphases: Exa for semantic search over quality sources, Tavily for agent-shaped search results, Firecrawl for turning arbitrary pages into clean markdown. Most people need one. Perplexity is worth adding if you want cited answers rather than raw pages.
Every server you add is attack surface
An MCP server runs with whatever credentials you hand it and can be prompted through the model. Prefer official servers over community forks with similar names, scope tokens to the minimum, and be deliberate about anything with write access to production. The convenience of an agent that can deploy is also the risk of an agent that can deploy.
A reasonable starting set
- Context7 — stops the hallucinated APIs
- GitHub — issues, PRs, code search
- Filesystem — local file access
- Playwright — browser testing
- Your database — Supabase, Neon, Postgres, whichever you run
- One browser server — Chrome DevTools to debug, Playwright to test
- One search server — Firecrawl, Exa or Tavily, not all three
That is six. Add the seventh when you feel an actual gap — Sentry for production errors, Browserbase when automation goes to production — not because a listicle recommended it.
If you use Claude specifically, the setup details differ slightly — see best MCP servers for Claude, and what MCP actually is if the protocol itself is new to you. The full MCP directory stays current.
Find these on the Radar
Every tool here lives on Kapyn Radar. Save the ones that fit into a Loadout and find them again.
