What is MCP? MCP stands for Model Context Protocol. It is an open-source standard for connecting AI applications to outside systems. Claude MCP is Claude using this standard to connect to data sources, tools, and workflows. Below, you’ll see what an MCP server is and how Claude connects to one.
What Is MCP: A USB-C Port for AI
If you want AI to check your calendar, look through a database, or operate other software, it first needs to be connected to the right tool or service. Wiring up every piece of software one by one would be a hassle. MCP sets common rules first, so anything built to those rules can work together.
The official docs use this analogy: MCP is like a USB-C port for AI applications. USB-C lets all kinds of electronic devices connect in the same way, and MCP lets AI connect to outside systems in the same way.
Examples the official docs give:
- An AI assistant reads your Google Calendar and Notion
- Claude Code builds a web app from a Figma design
MCP isn’t only for Claude. ChatGPT, Visual Studio Code, Cursor, and other tools support it too. If you first want to understand how AI does things on its own, see the complete AI Agent guide.
What Is an MCP Server: The Device You Plug In
MCP has three roles. An AI application like Claude is the host. It connects through an MCP client inside it to an MCP server. The server provides tools and data.
Sticking with the USB-C analogy: MCP is the port, and the MCP server is the device you plug in, such as a printer. Once it’s connected, the AI gains the features that device offers. In the official docs’ example, once you connect a Notion server, the AI can answer “what’s in my Projects database.”
Local MCP vs. Remote MCP
An MCP server can live in different places.
- Local MCP server: runs on your own computer.
- Remote MCP server: an MCP server you reach over the internet, which provides remote services and tools.
The remote MCP servers listed in Anthropic’s docs are third-party services. Anthropic doesn’t own, operate, or endorse them. The docs advise connecting only to servers you trust, and reading a server’s security practices and terms before you connect.
How Claude MCP Connects: The MCP Connector
The MCP connector is a feature of the Claude API. That means it is for developers who write code to connect to Claude, not a button in the Claude chat screen. It lets developers connect to remote MCP servers straight from the Messages API, without writing a separate MCP client. A few key points:
- You can turn on all tools, turn on only specific tools, or block tools you don’t want
- For servers that require sign-in, it supports connecting with an authorization token
- One request can connect to multiple servers
To connect, read the server’s documentation, get your authentication credentials ready, and then follow each company’s connection instructions.
Once connected, Claude calls a tool only when your request matches what the tool can do. Claude also has a developer service called Managed Agents. Its MCP tools require the user’s approval before every call by default. For more on using Claude, go back to the complete Claude guide.
Limits to Know Before You Use It
- It’s still in beta. The Messages API MCP connector is labeled beta, and developers have to turn it on separately in their code.
- It only reaches public servers. It requires the server to be public on the internet, so it can’t reach a local server on your computer. Managed Agents can reach private servers through MCP tunnels, but that is a research preview and you have to apply first.
- It only supports letting AI call tools. The MCP spec has many features, and the connector currently supports just this one.
- Think before you authorize. You have to judge for yourself the risk of the data and actions you give a server. Before handling sensitive regulated data, such as medical data, check that doing so complies with the rules.
This is based on the official docs as of November 9, 2026. Beta rules may change, so check the docs again before you use it.