Skip to content
AI Dev Toolkit.
Esc
  • AI Token CounterCount tokens for GPT, Claude, Gemini, DeepSeek, Qwen and more.Tool
  • LLM API Cost CalculatorEstimate per-request, daily and monthly API costs.Tool
  • AI Model ComparisonCompare prices, context windows and features across models.Tool
  • AI Model Pricing PagesSpecs, real costs and cheaper alternatives for popular models.Tool
  • Context Window CheckerSee whether your text fits each model’s context window.Tool
  • Subscription vs API CalculatorFind out whether a chat plan or the API is cheaper for you.Tool
  • GPU / VRAM CalculatorCheck how much VRAM a local model needs and which GPUs fit.Tool
  • Prompt Caching CalculatorEstimate savings from prompt caching.Tool

Coding Agents

MCP servers list: find the best MCP servers

Search 20,687 MCP servers from the Official MCP Registry, or start with the 58 we reviewed by hand, each with install snippets for Claude Code, Claude Desktop, Cursor, VS Code and Devin Desktop. Free, and the search runs in your browser.

58 reviewed servers

Steps

How to use the MCP server directory

  1. Search the reviewed servers by what you want the assistant to do, such as “postgres” or “browser”.
  2. Switch to “All registry servers” to search all 20,687, filtered by local or remote and by package type.
  3. Open a reviewed server to read what it can do, its licence and its security notes.
  4. Copy the snippet for your app, or open the server in the MCP config generator to combine it with others.

Method

How it works

An MCP server gives an AI app such as Claude Code, Cursor or VS Code a set of tools: read a GitHub issue, query a database, click through a web page. There are thousands of them, and the hard part isn’t finding one but choosing one you can trust. This directory has two layers. The full list comes from the Official MCP Registry, refreshed daily, minus the entries filtered out below. On top of it sit 58 servers we reviewed by hand on 2026-10-14: we read each README and repository, confirmed the package or endpoint is current, and wrote the description and install steps ourselves.

How to choose an MCP server

  • Prefer the official one. If the company behind the service publishes a server, use it. It follows their API changes and their sign-in. Community servers for the same service are often older copies.
  • Prefer remote where it exists. A remote server needs nothing installed and can’t read your files. 44% of the servers listed here offer a URL; 40% ship an npm package.
  • Check it’s alive. Look at the last update date and the repository’s recent commits. Several popular early servers, including the original GitHub, PostgreSQL and Slack reference servers, are archived but still appear in older lists.
  • Keep the list short. Each connected server adds at least its tool names and instructions to the model’s context, and many apps send every tool definition with every request, so servers you rarely use still cost tokens. The context window checker shows how much room you have.

MCP servers run code with your permissions

A local server is an ordinary program started by your AI app. It can read any file your user account can, use your network and run commands. The registry confirms that a publisher owns the GitHub account or domain in a server’s name, but its moderation policy says to assume “minimal-to-no moderation”: it removes malware and spam when it finds them, not buggy or vulnerable servers. The MCP project’s local server security guide recommends checking who publishes a package and its exact name (typosquatting is common), pinning a version, giving each server its own narrowly scoped token, limiting file servers to one folder, and running third-party servers in a container. Tool descriptions and tool results are also untrusted input: a web page or issue the assistant reads can contain instructions aimed at it, so keep approval prompts on for anything that writes.

How to install a server in each app

Every app reads MCP servers from its own file. Claude Code uses .mcp.json in the project or claude mcp add; Claude Desktop uses claude_desktop_config.json for local servers and Connectors for remote ones; Cursor uses .cursor/mcp.json; VS Code uses .vscode/mcp.json with a servers key; and Devin Desktop (formerly Windsurf) uses mcp_config.json. Each reviewed server’s page shows all of them, generated by the same code as the MCP config generator, so secrets are always written as variables, never pasted into the file.

Where the data comes from

Registry data is published under CC0. We keep only what the list shows: name, description, version, update date, source and website links, and package or remote details. Registry text is shown as plain text, links must use https, and package names must match their registry’s rules. Servers the registry marks deleted or deprecated are dropped. We don’t show logos. This directory is independent: it isn’t affiliated with, endorsed or sponsored by the Official MCP Registry, the Model Context Protocol project or any server’s maintainer.

The registry is open to anyone who can prove they own a GitHub account or domain, and a few publishers flood it: on 2026-10-15, 9,445 entries came from namespaces that publish more than 25 servers each. So the list keeps a server only when it links a source repository you can read (12,259 don’t), its namespace publishes 25 servers or fewer, and its description isn’t a placeholder shared by five or more entries (62). The full list is about 1340 KB, so it only downloads when you switch to it.

Examples

Reviewed MCP servers by category

Reference servers

Browser automation

Docs and web search

Code and developer tools

Projects and team work

Cloud and DevOps

Databases

Data and analytics

Payments and commerce

AI and machine learning

FAQ

Frequently asked questions

What are the best MCP servers?

It depends on your work, and a short list beats a long one, because every server’s tools take up room in the model’s context. For coding, most people start with a browser server (Playwright), a docs server (Context7) and the official server for their code host, such as GitHub. Then add the official server for each service you use daily, like your database or issue tracker.

Where does this list of MCP servers come from?

From the Official MCP Registry, the public catalogue run by the Model Context Protocol project. We read its API every day and list the latest active version of each server that links its source code: 20,687 on 2026-10-15. We leave out 21,766 more: servers with no source repository, namespaces that publish more than 25 servers each, and placeholder descriptions. The reviewed servers are our own picks, checked by hand.

Is every MCP server in the registry safe?

No. The registry checks that a publisher owns the GitHub account or domain in a server’s name, but its moderation policy says to assume minimal moderation, and it doesn’t remove buggy or vulnerable servers. Treat an unreviewed server like any package from the internet: read its source, check the publisher, and pin a version.

What’s the difference between a local and a remote MCP server?

A local server is a program your app starts on your computer, usually with npx, uvx or Docker, and talks to over stdio. It runs with your permissions. A remote server runs on the provider’s side and your app connects to its URL, often signing in with OAuth. Remote servers need nothing installed, so they’re the safer choice when one exists.

How do I install an MCP server in Claude Code?

For a remote server run claude mcp add --transport http <name> <url>, and for a local one claude mcp add <name> -- <command> <args>. Each reviewed server’s page has the exact command, plus the config for Claude Desktop, Cursor, VS Code and Devin Desktop. To set up several at once, use the MCP config generator.

Why don’t all servers have a page?

A page only helps if it says something the registry doesn’t. Reviewed servers get one with our own description, tested install snippets, the licence and security notes. The rest are listed with the registry’s own one-line description and a link to their source, so you can judge them yourself.

Why does “Configure by hand” not fill in the generator for unreviewed servers?

Because a link that pre-fills a command could be used to trick someone into running a package nobody has checked. Reviewed servers open in the generator ready to go, from our own data. For others, copy the package or URL from the server’s source and add it as a custom server.

Related

Related tools

Read

Guides

Glossary

Key terms