الأدوات: الإجراءات التي يتخذها النموذج

MCP gives a model three building blocks. The first is the tool: a function the AI can actually run to do something in the world. 🔧

A tool is for doing, not just knowing. Sending an email, querying a database, or creating a file all belong here because each one changes or fetches live state.

You expose the tool, but the model chooses when to call it. Based on the conversation, the AI picks the right tool and fills in its arguments for you.

In the Python SDK a tool is a normal function. You mark it with the @mcp.tool() decorator and the server advertises it to clients.

Each tool carries a name the model uses to call it. By default the function name becomes the tool name, so add and get_weather just work.

The model reads your docstring to learn what a tool does. A clear one sentence summary makes the AI pick the right tool at the right moment.

Your function's parameters turn into an input schema. The model sees the names and types so it knows exactly what data to send.

Whatever your tool returns goes back to the model as context. The AI then weaves that result into its reply to the user.

Because a tool runs real code, it can change things: write to disk, charge a card, or post a message. That power is exactly why tools sit under user control. ⚠️

MCP calls tools model-controlled. The model initiates the call, while the host app can still ask the user to approve it before anything runs.

Keep each tool focused on a single clear action. Small, well named tools are far easier for the model to choose correctly than one giant catch-all.

Let's lock in what makes a tool a tool.

Nice work! A tool is a model-controlled function that does something. You name it, describe it, type its arguments, and return a result the AI can use. ✅