What is an MCP server? A plain-English guide for Claude and other agents

An MCP server is a small program that lets an AI app use one outside tool or data source through the Model Context Protocol (MCP), an open standard Anthropic released in November 2024. The AI app, such as Claude, Claude Code, ChatGPT or Cursor, connects to the server, asks what it can do, and calls its tools when your request needs them. A GitHub MCP server lets Claude read your pull requests; a Stripe MCP server lets it look up a subscription.

Why MCP exists

Before MCP, every AI app needed its own integration for every tool: one plugin for GitHub in this app, another for GitHub in that one, and the same again for Jira, Slack and the rest. MCP replaces that with one shared plug shape. A company writes an MCP server once, and any app that speaks the protocol can use it.

That is why the same server URL now works in Claude, Claude Code, Cursor, VS Code and ChatGPT. In December 2025 Anthropic handed MCP to the Agentic AI Foundation, a new foundation under the Linux Foundation, so no single vendor controls it.

Hosts, clients and servers

The protocol names three parts. You only ever add the server; the other two come with the app you already use.

A host with five servers runs five clients, each with its own connection. Messages between them use JSON-RPC 2.0, so a server written in Python works with a host written in TypeScript, and the server never needs to know which model is on the other side.

What a server offers: tools, resources and prompts

A server can expose three kinds of things:

• Tools are actions the model can call, each with a name, a description and an input schema. list_pull_requests, create_issue and run_query are typical. The model decides when to call one, and hosts such as Claude Code ask you before a tool runs for the first time.
• Resources are data the app can pull into the conversation: a file, a database schema, an issue. In Claude Code you attach one with an @ mention, like @github:issue://123.
• Prompts are reusable templates the server ships, which hosts usually offer as commands in the / menu.

A server can also stop and ask you something mid-task, such as a confirmation or a missing value. The protocol calls this elicitation.

In practice, tools do most of the work. When you ask "which PRs are waiting on my review?", the host sends your question and the list of available tools to the model. The model picks the GitHub server's pull request tool with a review filter, the host routes that call to the server, and the result comes back into the answer.

Local and remote MCP servers

Servers run in one of two places, and the difference shapes setup and security.

Many SaaS vendors chose remote servers: there is nothing to install, you sign in with your normal account, and the vendor keeps the server current. Local servers still make sense for anything that has to touch your machine, like a browser the agent drives or a database on localhost.

MCP connectors in Claude

In the Claude apps, a remote MCP server is called a connector. As of October 2026:

• You browse reviewed connectors in the Anthropic Directory, or add any remote server by URL under Customize > Connectors > Add custom connector. Free plans are limited to one custom connector.
• On Team and Enterprise plans, an owner adds a connector for the organization first, then each person signs in with their own account and permissions.
• You switch connectors on or off per conversation from the + menu in the chat box.
• If you sign in to Claude Code with the same claude.ai account, those connectors load in Claude Code too.

So "MCP connector" and "MCP server" describe one thing from two sides: the server is the program, and the connector is the entry in Claude that points at it.

How to add an MCP server to Claude Code

In Claude Code, one terminal command adds a server. For a remote one:

claude mcp add --transport http linear https://mcp.linear.app/mcp

Then start claude, type /mcp, select the server and choose Authenticate to sign in. For a local server, everything after -- is the command that starts it:

claude mcp add playwright -- npx -y @playwright/mcp@latest

claude mcp list shows every server and whether it connected. Scopes, the .mcp.json file, tokens and fixes for common errors are in the Claude Code MCP setup guide.

Examples of MCP servers

Many of the apps teams use every day publish an official server. A few, with what you could ask once one is connected:

Each setup page lists the official server, the API route for when you would rather not use one, and the narrowest permissions that still work.

Security basics

An MCP server acts with whatever access you give it, and the model decides when to call it. A few habits keep that safe:

• Install servers from the vendor or a source you trust. A server is code, and a local one runs with your user's permissions.
• Give it the narrowest access that works. Read-only scopes, a restricted Stripe key instead of the secret key, a fine-grained GitHub token limited to the repositories in question.
• Read tool calls before you approve them. Keep "always allow" for read-only tools from servers you know.
• Watch for prompt injection. Text a tool returns, such as an issue body, an email or a web page, can contain instructions aimed at the model. A server that reads untrusted content and can also write or send things deserves extra care.
• Keep tokens out of shared files. In a project's .mcp.json, write ${GITHUB_TOKEN} instead of the token itself, so the file can be committed safely.
• Remove servers you no longer use. Fewer tools means less that can go wrong, and less of the context window spent on tool definitions.

MCP servers and skills

An MCP server gives an agent access, but it does not tell it what job to do. That is what a skill is for: a short SKILL.md with the steps for one task, like "every Monday, pull last week's MRR from Stripe and explain what changed". The two work as a pair, and skills vs MCP covers when you need which.

The skills library has ready-made skills that use these servers on a schedule and post the answer to your Briefkin board as one card, so you get the result without opening the app.