What is tool use (function calling)?

Ask a chatbot what the weather looks like for the weekend, and you may get a real forecast, with temperatures and the chance of rain. That's a little odd when you think about it. A language model learned from text that was collected months ago, and it can't look out of a window. So where did the forecast come from? The model asked for it. The app it lives in ran a weather lookup and handed back the result. That handoff is called tool use, and developers also call it function calling.

On its own, an artificial intelligence (AI) model can only read text and write text back. That leaves gaps people notice quickly.

• It doesn't know today's news, prices, or weather.
• It can get exact arithmetic wrong, because it predicts digits like it predicts words.
• It can't see your calendar, your files, or your email.

Tools fill those gaps. A search, a calculator, or a calendar lookup can each do what the model can't, as long as something runs them.

So here's the loop. The model reads your question and decides a tool would help. It writes a short, structured request: which tool, and the details it needs, such as the city for a forecast. The app runs that tool, then passes the result back. Finally the model writes its answer using what came back. Anthropic's developer guide, from the company that makes the Claude models, describes it that way: the model returns a structured call, and the application executes it.¹

How does the model know which tools exist? The app tells it, ahead of time, with a short menu. Each tool has a name and a plain description of what it does and when to use it. The model picks from that menu and nothing else. Anthropic's guide calls detailed descriptions by far the most important factor in how well tool use works. A vague one can lead the model to pick the wrong tool, or use the right one badly.¹²

The part people often miss is who does the work. The model never runs the tool itself. It writes a request, and the software around it decides whether to carry it out. That split matters. Because the app sits in the middle, the people who build it can choose which tools to offer, check a request before running it, and refuse ones that don't make sense.¹

A weather lookup only reads. Other tools can act: send an email, add a meeting, move money, or delete a file. With those, a wrong request has real effects, so that's where limits and approvals belong. A careful app asks you before a step you can't undo. Topic 60, on keeping a person in the loop, covers those pauses.

Tools are also what make agents possible. An agent is a model that works toward a goal in steps, and each step is often a tool call followed by a look at the result. Many tools now connect through a shared standard, the Model Context Protocol (MCP), so one tool can work with many apps. The next time an assistant shows a note like "searching" or "checking your calendar," you're watching a tool call. Which tools does your assistant have switched on?