A searchable, filterable directory of Model Context Protocol (MCP) servers — the small JSON-RPC processes that let Claude Desktop, Cursor, VSCode, Zed and Cline call out to your filesystem, databases, APIs and SaaS tools. Includes every official Anthropic reference server plus the most popular community implementations across databases, vector stores, cloud platforms, productivity apps, browser automation and more. Copy the install command or jump straight into the MCP Config Builder.
A Model Context Protocol (MCP) server is a small program — usually written in Python or Node.js — that exposes one or more tools, resources or prompts to an LLM client over a standard JSON-RPC interface. Once a client like Claude Desktop, Cursor or VSCode launches an MCP server (via its config file), the LLM gains the ability to call those tools directly: list files in a directory, query a database, fetch a webpage, post a Slack message, search a vector store, and so on. MCP is the closest thing the LLM ecosystem has to USB — one wire protocol, many possible peripherals.
The protocol was introduced by Anthropic in November 2024 and has been adopted in less than a year by Cursor, Continue, Cline, Goose, Zed, Windsurf and a growing list of other AI-first applications. The combination of a simple stdio transport, a clean JSON Schema for tool arguments and Anthropic's reference SDKs has made it easy for individual developers and companies to publish servers — which is why this MCP servers directory already lists more than sixty of them, with new ones appearing every week.
Use the search field at the top to find a server by name or by what it does ("notion", "vector", "postgres", "browser"). Use the category chips to narrow by domain: official Anthropic reference, databases, vector stores, cloud and infrastructure, productivity, communication, web scraping, AI utilities. Each card shows:
node, python or docker — so you know which package manager will fetch it.npx or uvx) you can paste into your terminal or your config file.claude_desktop_config.json; View source links to the GitHub repo.If this is your first MCP server, start with Filesystem from the official set. It's the easiest to verify — point it at a folder, and Claude immediately gains the ability to list, read and write files in that folder. From there, the most rewarding next steps depend on what you actually do all day:
Every MCP server in this directory runs locally on your computer. The personal access tokens, API keys and connection strings you put in your config file are read by the server process — not uploaded to Anthropic, not uploaded to Tools Searcher. Even when Claude calls a tool, only the tool's response (which the server fetches on your behalf using your credentials) is sent to the model. Treat the config file like any other secrets file: keep it out of git, prefer scoped tokens, rotate them periodically.
This directory covers the most-used servers as of mid-2026, but new ones ship every week. If you don't see a server you want, two things to try. First, the MCP Config Builder has a Custom server tab where you can plug in any command, args and env — useful for anything you build yourself or fetch from a personal fork. Second, if you've authored a server you'd like added, the official modelcontextprotocol/servers repo on GitHub and the community-maintained awesome-mcp-servers list are the main discovery channels we monitor.
npx -y @modelcontextprotocol/server-filesystem /Users/you/DocumentsGITHUB_PERSONAL_ACCESS_TOKEN=ghp_xxx \
npx -y @modelcontextprotocol/server-githubuvx mcp-server-fetchMCP is an open standard from Anthropic (introduced November 2024) for connecting LLM clients to external tools, resources and prompts. An MCP server is a small process that exposes capabilities; an MCP client (Claude Desktop, Cursor, VSCode, Zed, Cline, Goose, Windsurf, etc.) launches the server, lists what it can do, and routes LLM tool calls to it over JSON-RPC.
The official Filesystem server is the most forgiving way to verify your setup — install it, point it at a single directory, restart your client, and Claude will immediately list and read those files. Once it works, layer in GitHub (if you code), Postgres or SQLite (if you query data), or Notion/Linear (if you live in those apps).
The official Anthropic servers are open source and audited by the community. Third-party servers vary — read the README and skim the code before installing any server that touches sensitive data. Treat MCP servers like any other npm or pip package: prefer well-known publishers, pin versions when stability matters, and grant credentials with the smallest possible scope.
To use a server, no — you only need to install it (via npx or uvx) and add a small JSON entry to your client's config file. The MCP Config Builder generates that entry for you. To write a server, you'll use the official Python or TypeScript SDK, which hides JSON-RPC behind decorators like @mcp.tool().
As of mid-2026: Claude Desktop, Cursor, VSCode (native MCP since 1.95), Zed, Cline (VSCode extension), Continue, Goose, Windsurf and a growing list of community clients. The protocol is standard, so any compliant client can run any compliant server — you don't need a separate server build for each client.
stdio is for local servers: the client launches the server as a subprocess and they talk over stdin/stdout. This is what every server in this directory uses by default. SSE/HTTP is for remote servers: the client connects to a URL, and the server runs somewhere else (a container, a cloud function). Remote MCP is still maturing — most production setups today are stdio.
Each card has a View source link that goes to the upstream GitHub repository. Official Anthropic servers live in github.com/modelcontextprotocol/servers; community servers are scattered across individual repos. Always glance at the README and recent commits before installing.
Yes — the recommended setup. Add as many mcpServers entries as you need. Each runs as its own subprocess and exposes its own tools; Claude (or another client) sees the union of all available tools across all servers. Pair this directory with the MCP Config Builder to stack them cleanly.