What Is an MCP Server? A Plain Explanation With an Example

An MCP server is a program that gives an AI assistant tools to call and data to read through the Model Context Protocol (MCP), an open standard for connecting AI applications to external systems. Instead of you copying information into a chat, the assistant asks the server for it. This guide explains the term, how a server differs from a client and a host, shows one worked example, and covers how to connect one safely.

MCP in one paragraph

MCP, the Model Context Protocol, is an open-source standard for connecting AI applications to external systems, and Anthropic introduced it in November 2024. The official docs compare it to a USB-C port: one standard way to plug an AI application into files, databases, search engines and other services, rather than a custom integration for each. In that picture, the server is the plug on the service side. The protocol only covers how context is exchanged. Per the architecture overview, it does not dictate how an AI application uses a language model or manages the context it receives.

What a server does vs a client (and host)

An MCP server provides context, a client maintains the connection to one server, and a host is the AI application that runs the clients. The architecture overview defines all three, and the host creates one client per server.

Role	What it is	Example
Host	The AI application that coordinates one or more clients	Claude Code, Claude Desktop, Visual Studio Code
Client	A component inside the host that holds the connection to one server	The connection object Visual Studio Code creates for each server
Server	A program that provides context to clients	The filesystem server, the Sentry server, LinkIntel

In day-to-day use you only meet the host and the server. The client is plumbing.

A server offers up to three kinds of things, which the spec calls primitives:

Tools: functions the assistant can run, such as a file operation, an API call or a database query.
Resources: data the assistant can read, such as file contents or database records.
Prompts: reusable templates, such as system prompts or few-shot examples, that structure how the assistant works with the language model.

A server can run on your machine or elsewhere. The two transports in the docs are stdio, where the host launches a local process and talks to it over standard input and output, and Streamable HTTP, which uses HTTP POST and lets a server be remote and serve many clients.

Example: asking Claude about your own X post results

LinkIntel is a hosted MCP server that lets Claude read how your own recent X posts performed, so a drafting suggestion can cite your numbers. It is a small, real example of a tools-only server, and it is deliberately narrow.

Here is what it does, and nothing more:

Once you connect X, it reads your recent original posts and their metrics every day and stores snapshots.
It reads X's open-source ranking weights every day, so the assistant can cite what the ranker is documented to reward.
Your assistant gets both as a cited content brief. A typical prompt is "Which of my last 20 posts did best, and what should I write next?"
It never publishes, and it never searches X.
It costs $39 USD a month.

The assistant calls a tool, LinkIntel answers from its stored data, and Claude writes the reply with the figures and their dates in front of it. For a deeper look at the X-specific version of the question, read what an MCP server for X post analytics is, and for setup steps see how to give Claude your X post data.

What an MCP server can and cannot do

An MCP server can only offer what its author built, and the assistant decides when to call it. That limits what a server can do in two ways.

It can:

Return data the assistant could not otherwise see, such as your files, your database or your own analytics.
Perform actions through write tools, if its author included them.
Announce when its list of tools changes, so the host can refresh.

It cannot:

Make the assistant smarter. It supplies context, and the model still decides what to do with it.
Act outside its own tools. A read-only server has no way to change anything, however a prompt is worded.
Guarantee a correct answer. The assistant can still misread good data, which is why servers that cite sources are easier to check.

A server also does not search the web or the rest of the internet unless it was built to. Ask what a given server touches before assuming.

How to connect one

You connect a remote MCP server by giving your host its URL, and a local one by telling the host which command to run. The steps differ slightly by host.

Claude Code. Run claude mcp add with the transport and the URL. For LinkIntel it is claude mcp add -s user --transport http linkintel https://www.getlinkintel.com/api/mcp.
Claude.ai or Claude Desktop. Open Settings, then Connectors, then Add custom connector, and paste the server URL.
Any other MCP-compatible host. Look for an MCP or connectors section in its settings. Remote servers take a URL, and local servers take a command and arguments in a config file.

Most hosted servers then open a browser window for sign-in. The LinkIntel MCP docs list its endpoint, tools and auth method, and the FAQ covers the rest of setup.

Security and what data leaves your machine

What leaves your machine depends on the server's transport and tools: a local stdio server runs on your computer, while a remote server receives whatever the assistant sends it. Treat a connection as giving a third party a defined slice of access, then check four things.

Authentication. For remote servers, the docs say the Streamable HTTP transport supports bearer tokens, API keys and custom headers, and that MCP recommends OAuth for obtaining tokens. OAuth in your browser beats pasting a long-lived secret into a config file.
Read-only or write tools. Look at the tool list. Any tool that creates, edits, deletes or publishes is something the assistant can call.
What is stored. A remote server may keep what it reads. Look for a privacy policy that says what is kept and how to delete it.
A server card, if the server publishes one. Not all do; when it exists, read it before you connect. LinkIntel publishes its server card, and it uses OAuth 2.1 with Google sign-in. It never asks for your X password, and disconnecting X deletes the stored snapshots.

If a server asks for a password or secret typed into a chat, stop.

Other examples of MCP servers

Several well-known servers show the range, from local file access to hosted developer tools. These are examples, not a ranking, and each is verified against its own source.

Filesystem. A reference server for secure file operations with configurable access controls, in the official servers repository. It runs locally over stdio.
Git. A reference server in the same repository for reading and working with Git repositories.
Fetch. Another reference server there, for retrieving web content for the assistant.
Sentry. The Sentry MCP server runs on Sentry's platform over Streamable HTTP, per the MCP architecture docs, which use it as the standard remote example.
GitHub. GitHub's official MCP server connects assistants to repositories, issues and pull requests, with a remote version hosted by GitHub.

The reference servers in that repository are described as educational examples rather than production-ready solutions. Older servers for GitHub, PostgreSQL and Slack in that project were moved to an archive, so check which implementation is maintained before you rely on one.

FAQ

What does an MCP server do?

An MCP server gives an AI assistant tools it can call, data it can read and prompt templates it can use. The assistant asks the server instead of you pasting the information into the chat.

What is the difference between an MCP server and an API?

An API is built for programmers, while an MCP server describes its tools in a standard format so an assistant can discover and call them on its own. Many MCP servers wrap an existing API.

Is an MCP server safe?

It depends on the server. Check how it signs you in, whether it has write tools, what it stores, and who runs it. Prefer OAuth consent and read-only access.

Do I need to code to use an MCP server?

No. Connecting a hosted server is one command in Claude Code or a form in Claude.ai. Building your own server needs code, but using one usually does not.

Who created MCP?

Anthropic introduced the Model Context Protocol as an open standard in November 2024 and in December 2025 donated it to the Agentic AI Foundation, a fund under the Linux Foundation. Many assistants and developer tools beyond Claude support it.