What is an MCP server?
An MCP server is a small program that gives an AI assistant access to one specific tool or data source — your database, your files, a GitHub repository, a Slack workspace. It speaks the Model Context Protocol, an open standard introduced by Anthropic in late 2024, so any compatible AI client can use it without a custom integration.
Before MCP, every combination of AI app and external tool needed its own bespoke integration — a connector for Claude and Postgres, another for a different assistant and the same database, and so on. MCP replaces that with one protocol: build a server once, and every compatible client can talk to it.
What an MCP server exposes
The protocol defines three kinds of things a server can offer. Most servers you'll meet are built around the first.
- Tools
- Actions the model can choose to run — query a table, open a pull request, send a message. This is what most servers are built for, and it's also where the risk lives: a tool can write and delete, not just read.
- Resources
- Read-only context the model can pull in — a file, a document, the result of a lookup. Resources inform the model's answer without letting it change anything.
- Prompts
- Reusable prompt templates the server offers the client, so a server can ship a known-good way of asking for the thing it does well.
How it connects to your AI client
- Local (stdio)
- The client launches the server as a subprocess on your own machine and talks to it over standard input/output. Nothing leaves your computer, which is why local servers are the common choice for filesystem and database access.
- Remote (HTTP)
- The server runs somewhere else and the client reaches it over HTTP. Convenient for hosted services, but your requests — and any credentials — travel to a third party, so it's worth knowing who operates it.
What real MCP servers look like
Abstract definitions only go so far. These are real servers from our directory, each connecting an assistant to one specific system:
How it differs from things it resembles
- A plain API
- An API is built for a programmer who reads the docs and writes the calls. An MCP server wraps an API and describes itself to the model — what it can do, what arguments it takes — so the assistant can decide when to use it without anyone hard-coding the call.
- A plugin
- Plugins are tied to the product that defines them. MCP is an open standard, so one server works across every client that supports the protocol rather than only inside the app it was written for.
- A Claude Skill
- A skill is instructions — it teaches the model how to do something well with what it already has. A server is a connection — it gives the model access to something it otherwise couldn't reach at all. They solve different problems and are often used together.
How do you add one?
You don't install an MCP server the way you install an app. You add a short JSON entry to your AI client's config file naming the command that starts the server, then restart the client. Each client keeps that file in a different place, and most servers need an API key or path passed in as well.
Full step-by-step install guide →
We generate the exact config entry, with safe defaults, for each client:
Is it safe to add one?
Anyone can publish an MCP server, and a server you add can read files, run commands, and handle credentials on your behalf. That's the whole point of it — and the reason it's worth checking what a server actually requests before adding it, especially one from an author you don't recognise.
Every server in this directory carries a security score computed from repository signals — maintenance activity, permissions requested, authentication, licence. What to check before installing one goes through the real risks, and the methodology shows how the score is calculated.
Frequently asked questions
›How does an MCP server work?
The AI client starts the server and asks it what it can do. The server replies with a list of its tools and what each one needs. From then on, when your request calls for one of them, the model asks the client to run that tool, the client runs it and passes the result back, and the model uses the answer in its reply. The model never touches your data directly — it only asks, and the server decides what to return.
›What does an MCP server actually do?
It translates. On one side it speaks whatever the underlying system speaks — SQL for a database, the GitHub API for a repository, the filesystem for local files. On the other it describes those capabilities in the format the Model Context Protocol defines, so any compatible AI client understands them without being taught about that specific system.
›Are MCP servers only for Claude?
No. MCP is an open standard, and the same server works in any client that implements it — Claude Desktop, Cursor, VS Code and Windsurf among them. A server written for one is not rewritten for the others; only the client's config file entry differs.
›Do I need to know how to code to use one?
To use an existing server, no — you add a few lines of JSON to your client's config file and restart it. Writing your own server does take programming, since a server is a real program that speaks the protocol.
›Are MCP servers safe to install?
It depends entirely on the server. Anyone can publish one, and an MCP server runs with whatever access you give it — files, credentials, the ability to run commands. Prefer servers from the vendor whose product they connect to, check what permissions the server asks for, and be more careful with one that can write or delete than one that only reads.