AI Explained
How a model uses tools
Tool use lets a model act. The Model Context Protocol makes every tool fit every AI application.
Tool use
How a model does things
A model by itself only writes text. Tools are how it looks things up and acts.
Press Play to walk through it, or Listen to have it read to you.
Read all the steps
- Your application sends the request together with a list of tools: each has a name, a description and the inputs it takes.
- The model does not run anything. It replies that it wants a tool, and with which inputs.
- Your code runs the tool.
- The result goes back to the model in a follow-up request.
- The model writes the answer, using the result.
The model asks, your code acts
Tool use lets the model call functions you define. It decides when a tool is needed from the request and the tool's description, and returns a structured call. Your application runs it [1].
What you describe
Each tool has a name, a description and a schema for its inputs, sent along with the request [1].
Two kinds
Client tools run in your application. Server tools, such as web search or code execution, run on the model provider's side and you only see the results [1].
MCP
One standard plug for every tool
Without a standard, every AI application needs its own connector to every system. The Model Context Protocol lets a connector be written once.
Press Play to walk through it, or Listen to have it read to you.
Read all the steps
- The host is the AI application. It creates one client for each server it connects to.
- A local server runs on your own machine and talks over standard input and output.
- A remote server runs somewhere else and talks over HTTP, usually for many clients.
- The client first asks the server what it supports, then asks for its list of tools.
- Servers offer three things: tools the model can call, resources it can read, and prompts that are ready-made templates.
- When the model wants a tool, the host sends the call through that server's client and passes the result back.
What it is
An open-source standard for connecting AI applications to outside systems. Its own comparison: like a USB-C port for AI applications [2].
Three participants
The host is the AI application. A client keeps the connection to one server. A server is a program that provides context to clients. The host creates one client for each server [3].
Local and remote
A local server talks over standard input and output and usually serves one client. A remote server talks over HTTP and usually serves many [3].
What a server offers
Tools are functions the application can call to act. Resources are data that give context. Prompts are reusable templates [3].
How a call goes
In the documentation's example the client asks the server what it supports, lists its tools, then calls one. The messages are JSON-RPC: server/discover, tools/list, tools/call [3].
What a server can ask back
Through elicitation, a server can ask the user for more information [3].
Build one
Write an MCP server in a minute
Name your server and describe its tools. The starter writes the Python file in the shape the official guide uses [4]. You fill in what each tool does.
The code follows the official Python guide read on the date below: the MCPServer class, a decorator on each tool, and the stdio transport. That guide warns never to print to standard output in a stdio server, because it corrupts the messages, so the starter logs instead. Check the guide for the current SDK version before you ship.
Sources
Where this comes from
Read 2026-11-15, unless a source gives its own date. A number in brackets points to the page a statement was read on. Sentences with no number, the small demonstrations and the tools are this center's own explanations.