mcpdirectory.dev
9,129 servers · updated 23 Sept

What is an MCP server?

An MCP server is a small program that gives an AI assistant the ability to do one specific thing — read a database, open a web page, search your notes. On its own an assistant can only talk. With a server attached it can act.

The short version

MCP stands for Model Context Protocol. It is an open standard that defines how an assistant like Claude or Cursor asks an outside system to do something, and how that system answers. Anthropic published it and it is now governed under the Linux Foundation, which is why servers built for one assistant tend to work in the others.

A server is not a machine you rent. In almost every case it is a package that runs on your own computer when your editor starts it, and stops when you close the editor. The word "server" is about the role it plays in the conversation, not about hosting.

What that looks like in practice

Without a server, asking your assistant "why did checkout fail last night?" gets you a guess. With a logging server attached, it reads the actual logs and tells you. The assistant does not become smarter — it gains reach.

Here is one server from each of the main categories, so the idea is concrete:

ServerCategoryLets an assistant
Supabase Databases Query and retrieve data from your Supabase database through an AI assistant.
Saidsef Mcp Github Pr Issue Analyser Code and repositories Fetch a pull request's diff and CI status, then post a review comment from an LLM.
Receipt Extraction Documents and notes Automate receipt processing for Australian businesses by uploading images or PDFs and receiving structured JSON with validated ABN, line-item tax codes, and GST calculations.
Browser Use Browser automation Fill out web forms and applications automatically with provided information.
Mockserver Communication Mock external APIs during development when those services are unavailable or incomplete, then verify your application calls them correctly.
Scanmalware Observability Submit URLs for malware scanning and retrieve risk scores, indicators, and AI analysis results

Do I need to write code?

No. Installing a server means pasting a short block of configuration into a file your editor already reads. Nothing is compiled, nothing is written from scratch. If you can edit a text file you can do this.

Try one now

This takes about two minutes. Pick your editor, pick something to connect, and paste the result. Start with a server marked no key needed — those work without you signing up for anything, which removes the step where most people give up.

  1. 1 · Which editor do you use?

  2. 2 · Pick something to connect

    These are working today and need the least setup. The ones marked no key needed run without you creating an account anywhere.

  3. 3 · Paste this in

    Open .mcp.json (in your project root) and add:

    Mockserver · Claude Code 
    {
  "mcpServers": {
    "mockserver": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "docker.io/mockserver/mockserver:7.6.0"
      ]
    }
  }
}

    Or from a terminal: claude mcp add mockserver -- docker run -i --rm docker.io/mockserver/mockserver:7.6.0

    Restart Claude Code, then ask it something only Mockserver would know. If it answers, you are connected. Full details for Mockserver

How to tell a good server from a bad one

Anyone can publish one. The official registry applies almost no moderation by design, and a large share of what is out there was written once and abandoned. Three things worth checking, in order:

  • Is it maintained? Every entry in this catalogue shows when it was last worked on. A server nobody has touched in a year will break against the next API change and nobody will fix it.
  • Is it the official one? When a vendor publishes their own, prefer it. Community forks exist for good reasons, but they are one person's spare time.
  • What can it reach? A server runs with whatever access you hand it. Point it at a development account before production.

Star counts are the one signal worth ignoring. Popularity records what was interesting a year ago, not what works today — which is why this catalogue sorts by maintenance instead. How that is calculated

Where to go next