What is an MCP server?
An MCP server is a small program that connects an AI application to one specific tool or source of information — your files, a database, GitHub, a calendar — using the Model Context Protocol (MCP). The AI application sends the server a structured request; the server does the job and returns the result. One server per tool, and any MCP-compatible AI app can use it.
This article is about that server component specifically. If you want the bigger picture — what the protocol is, who invented it and why it spread so quickly — start with our full protocol explainer. The short version: MCP is an open standard created by Anthropic in November 2024 and, since December 2025, stewarded by the Agentic AI Foundation under the Linux Foundation. The protocol is the shared language; the servers are the things that actually speak it on behalf of your tools.
The word "server" can mislead. An MCP server is not necessarily a big machine in a data centre — it is often a few hundred lines of code running quietly on your own laptop, launched and closed by the AI app itself. What makes it a "server" is its role: it waits for requests and serves responses.
The restaurant analogy
Picture the AI model as a customer in a food hall with many kitchens. The customer cannot walk into the kitchens and start cooking — and should not. Instead, each kitchen has its own waiter.
An MCP server is that waiter. It stands in front of one kitchen — the filesystem kitchen, the database kitchen, the GitHub kitchen — and offers a written menu of what can be ordered: "read a file", "run a query", "open a pull request". The customer places an order in a standard format, the waiter does whatever messy work the kitchen involves, and comes back with the dish. The customer never needs to know how the kitchen operates, only what is on the menu.
The Model Context Protocol is the ordering convention the whole food hall agrees on — the standard way menus are written and orders are placed. Because every waiter follows the same convention, any customer (Claude, ChatGPT, Cursor, VS Code) can order from any kitchen without learning a new system. That is the entire trick: the protocol standardises the conversation; the server does the work.
Host, client and server: who does what
The official MCP architecture documentation defines three participants, and the terms are worth keeping straight because they get muddled constantly online.
| Term | What it is | Everyday example |
|---|---|---|
| MCP host | The AI application you actually use, which coordinates everything | Claude Desktop, Claude Code, VS Code, Cursor, ChatGPT |
| MCP client | A component inside the host that maintains one dedicated connection to one server | Invisible to you — the host creates one per server |
| MCP server | The program that provides context and capabilities from a tool or data source | Filesystem server, GitHub server, PostgreSQL server |
So when you "add three MCP servers to Claude Desktop", the host quietly spins up three clients, each holding a dedicated line to its server. You never touch the clients; you only choose the servers. Casual writing often uses "MCP client" to mean the host app itself — harmless in conversation, but the spec keeps them separate.
What an MCP server actually does
Under the current MCP specification (version 2025-11-25, with the next revision due 28 July 2026), a server can offer three kinds of things to the AI:
- Tools — actions the model can ask to perform: create a file, run a database query, send a message, open a pull request. This is the headline feature and the reason most servers exist.
- Resources — data the model can read for context: file contents, database schemas, documents, log entries.
- Prompts — reusable templates the server provides for common workflows, which the user or host can invoke.
A concrete example: the GitHub MCP server exposes tools such as "list my open issues" and "create a pull request", plus the repository data the model needs to use them sensibly. Connect it to a host and your assistant can manage repositories in plain conversation — our GitHub MCP server setup guide for Claude Code walks through it step by step.
Two properties are worth underlining. A server is narrow by design — it fronts one system, and you combine several to build up capability. And it is passive — it never acts on its own initiative; it only answers requests, a point that matters for the agent question below.
How you run one: local vs remote
The specification defines two transport mechanisms — two ways the bytes travel — and they map neatly onto the two ways people run servers in practice.
Local, over stdio. The host launches the server as a small process on your own machine and talks to it through standard input and output. This suits anything touching local data: your files, a folder of documents, a database on your laptop. Credentials stay on your machine. Setup usually means adding a few lines to a configuration file (such as claude_desktop_config.json or an mcp.json) naming the server and how to start it. Our PostgreSQL MCP setup guide shows a complete worked example of a local server.
Remote, over Streamable HTTP. The server runs on someone else's infrastructure — GitHub's, Sentry's, your company's — and the host connects across the internet, typically signing in with OAuth. This is what powers the one-click "connectors" in Claude and ChatGPT: under the surface, each connector is a remote MCP server. One remote server can serve thousands of users at once, whereas a local stdio server typically serves just one.
Host support is now broad: Claude's apps, ChatGPT (OpenAI adopted MCP in March 2025), Cursor, VS Code, Gemini-based tools and most agent frameworks all connect to the same servers.
MCP server vs MCP agent: not the same thing
Search data shows plenty of people asking "what is an MCP agent?", so let's untangle it. "MCP agent" is not a formal term in the specification — the spec defines hosts, clients and servers. When people say "MCP agent", they almost always mean an AI agent that uses MCP servers as its hands: a model, running inside a host application, that plans a task, calls tools on one or more servers, checks the results and keeps going.
The division of labour is clean. The agent is the active part — it decides what to do. The server is the passive part — it sits waiting and does precisely what it is asked, nothing more. A filesystem server will never spontaneously reorganise your documents; an agent connected to it might, if you ask. If a product markets itself as an "MCP agent", read that as "an agent that speaks MCP", not as a special kind of server.
Where to find MCP servers
The ecosystem has grown from a handful of demos in late 2024 to thousands of published servers. Sensible places to look, in order:
- The official MCP Registry — an open catalogue at registry.modelcontextprotocol.io, launched in September 2025 as the standard way to publish and discover servers.
- The reference servers repository — the steering group maintains a set of reference implementations (Filesystem, Fetch, Git, Memory and others); the repo has over 85,000 GitHub stars.
- Official vendor servers — GitHub, Sentry, Cloudflare, Stripe and many others ship their own, which are generally the safest choice for that service.
- Curated lists — for a ranked, plain-English shortlist, see our best MCP servers for 2026, and browse the wider AI infrastructure directory for the tooling around them.
Remember the shape of it: the protocol is the plug standard, hosts are the appliances, and servers are the adapters that make your tools fit the socket. For everything upstream of the socket, the full protocol explainer picks up the story.
Frequently asked questions
What is an MCP server?
An MCP server is a small program that connects an AI application to one specific tool or data source — such as your files, a database or GitHub — using the Model Context Protocol. The AI application sends it a structured request, the server does the work and returns the result.
What are MCP servers used for?
MCP servers give AI assistants and agents standardised access to outside systems: reading and writing files, querying databases, searching the web, managing code repositories, sending messages and more. Each server handles one tool or service, and you combine several to build up an assistant's capabilities.
What is an MCP agent?
"MCP agent" is an informal name for an AI agent that uses MCP servers to act on the world. The agent is the active part — an AI model inside a host application, deciding what to do next. MCP servers are the passive plumbing it calls on. The MCP specification itself defines hosts, clients and servers, not agents.
Do I need to be a developer to use an MCP server?
Not any more. Remote MCP servers surface as one-click connectors in apps such as Claude and ChatGPT and can be enabled from a settings menu with no code. Running a local server still involves editing a small configuration file, but step-by-step guides make it manageable for careful non-developers.