MCP Directory

The best MCP servers, by what you actually need

There is no single “best” MCP server — it depends on the job. For coding agents it's a filesystem server, the official GitHub server, and an up-to-date docs server like Context7. For data it's a guarded SQL server; for research, an AI-native search server plus fetch. Below are the top picks by task, capability, client and type — drawn from 470+ curated servers and judged on one thing: does the config actually work.

How we judge a server

“Best” here doesn't mean most stars. A server earns a place if it's worth wiring into a real agent. Four signals matter more than popularity:

Official, strictly

The repo belongs to the vendor — not a lookalike fork. First-party servers track the API and get security fixes.

Actually maintained

A recent commit/release. An abandoned server is a liability no matter how clever it was.

A sane tool budget

Each tool eats context. We show the real tool count so you don't blow your client's tool limit on one server.

A config that works

Exact mcpServers JSON for Claude, Cursor and Windsurf — transport and auth correct, not a guess.

See the full methodology and the live numbers in the State of MCP data report.

Best MCP servers by job

Pick the workflow closest to yours — each guide is an opinionated shortlist (what to add, and what to skip).

Best MCP servers for coding agents (Claude Code, Cursor, Codex)

For a coding agent, add four and stop: filesystem (let it read/write your repo), GitHub (issues/PRs), a docs-context server like Context7 (kills hallucinated APIs), and Playwright only if you actually test UIs. That's ~35 tools — right at Cursor's limit. Everything past that makes the agent pick worse, not better.

Best MCP servers for data analysis & SQL

Start with a read-only database server (Postgres or your warehouse) — that single guardrail matters more than which one you pick. Add a notebook server (Jupyter) if you want the agent to actually run analysis, not just query. The risk here isn't capability; it's an unguarded execute_query against production.

Best MCP servers for web search & research

For AI-native semantic search, Exa or Tavily. For turning messy pages into clean markdown the model can read, Firecrawl. For privacy-respecting general search, Brave. Most people only need one search server plus one fetch/scrape server — and should never pair web reading with broad write access in the same session.

Best MCP servers for DevOps & infrastructure

Pick the official server for your stack — GitHub for CI, Kubernetes for clusters, Terraform for IaC, your cloud's own server (AWS) for everything else. But the pick matters less than the discipline: never point an agent at prod, and gate anything irreversible (an apply, a delete, a deploy) behind explicit human approval.

Best MCP servers for notes, docs & knowledge

Use the official Notion server if you live in Notion; for local-first notes, an Obsidian server keeps everything on your machine. Add a memory server only if you genuinely need persistence across sessions — and keep your expectations modest; it's a key-value store, not understanding.

Best MCP servers by capability

Already know the one thing you need the agent to do? Jump straight to the servers that do it.

See all capabilities

By client

Same servers, the exact config for your client.

By type

Filter by how a server runs and who maintains it.

FAQ

What are the best MCP servers right now?

There's no single winner — the best MCP server depends on what your agent needs to do. For coding agents, the core set is a filesystem server, the official GitHub server, and an up-to-date docs server like Context7. For data work it's a guarded SQL server (Postgres or SQLite). For research it's an AI-native search server (Exa or Tavily) plus a fetch server. All of these are on mcptrove with a copy-paste config for Claude, Cursor and Windsurf.

What's the best MCP server for Claude Desktop or Cursor?

The same servers work across clients — MCP is a shared protocol. The difference is only the config shape, and every server page gives you the exact JSON for Claude Desktop, Cursor and Windsurf. Start with the filesystem and GitHub servers, then add one or two specific to your task.

Are official MCP servers better than community ones?

Usually, yes — an official (vendor-maintained) server tracks the vendor's API, gets security fixes, and won't quietly disappear. That's why mcptrove verifies "official" strictly: the repo must belong to the vendor. But a well-maintained community server is often the right call when no official one exists; check freshness and the tool list either way.

How many MCP servers are there?

The wider ecosystem has thousands of servers across registries, but most are duplicates, forks or abandoned. mcptrove curates 470+ where each one is real, deduped, and shown with its transport, auth, real tool list and a freshness date.

How is this different from an "awesome MCP servers" list?

An awesome-list is a flat list of links — you still click through to each repo and figure out setup yourself. This is a working catalogue: every server has a ready-to-paste config per client, the real tool list, a strict official flag, and capability and comparison pages so you can actually choose between them.