What is Model Context Protocol?

What it is

The Model Context Protocol, usually shortened to MCP, is an open standard for connecting AI applications to the tools and data they need. Anthropic introduced it in late 2024, and many other AI companies and developers have since adopted it.

Think of a USB port. Before USB, every device needed its own cable and its own driver. MCP does the same job for AI: instead of writing a custom connection between every AI app and every service, a developer builds one MCP connection and any app that speaks the protocol can use it.

How it works

MCP has two sides. An MCP server wraps something useful, such as a calendar, a database, a code repository or a company wiki, and describes what it can do. An MCP client lives inside an AI app, such as a chat assistant or a coding tool, and lets the model discover those abilities and call them.

A server can offer three kinds of things:

• Tools: actions the model can ask to run, like "create a ticket" or "run this query".
• Resources: data the model can read, like a file or a record.
• Prompts: reusable instructions a user can pick from.

When you ask a question, the model sees which tools are available, decides whether one would help, and asks the client to call it. The result comes back as text the model can read, which uses up part of its context window. The model never gets direct access to your systems. The client and server sit in between and apply the permissions you set.

Why it matters to you

MCP is what turns a chatbot into an AI agent that can actually do things in your software. It also reduces lock-in. If your tools speak MCP, you can switch the AI model or the app on top without rebuilding every integration.

Things to watch

Because an MCP server can act on your behalf, you should treat each one like an app you install. Connect only servers you trust, give them the narrowest access that works, and check what they can change. Text returned by a tool can also contain hidden instructions meant to trick the model, an attack called prompt injection, so sensitive actions deserve a human confirmation step.

An example

Say you connect a project-tracker MCP server to your assistant. You ask, "What is overdue on the launch board?" The assistant calls the server's search tool, reads the results and answers in plain English. Ask it to move a task and it calls a different tool, ideally after asking you first.

Related ideas

MCP sits next to tokens and context limits in practice, because every tool description and every result costs tokens. Teams that use many servers often watch that cost and use prompt caching to keep it down.