In plain words
On its own, a language model can only write text from what it learned in training. It cannot see today's weather, check your calendar or do exact maths on a huge number.
Tool use fixes this. The app tells the model which tools exist, such as "weather", "web search" or "send email", and what each one needs. When a question calls for one, the model writes a short, structured request, for example weather("Pune"). The app runs the real tool, hands the result back, and the model writes its answer with fresh facts.
The model never presses the button itself. It asks, the app does it. That split is what lets apps add approvals and limits. Tool use is the basic move behind AI agents.
Everyday example
A shop assistant on the phone with the godown
You ask a shop assistant whether a size 9 is in stock. She doesn't know offhand, so she calls the godown: "Size 9, black, model 42?" The godown checks and says "two left". She tells you: "Yes, we have two." She asked the right question in the right format, someone else checked, and she turned the answer into a reply.
Where the example breaks: The assistant knows when the godown sounds unsure. A model trusts whatever text comes back from a tool, so a wrong or tampered result flows straight into its answer.
How it actually works
In developer terms this is called function calling or tool calling. It is a loop between the model and the app.
Describe the tools
With each request, the app sends a list of tools: a name, a plain description of what it does, and the inputs it takes, written as a JSON Schema. These descriptions sit in the model's context window.
The model asks for a call
If a tool would help, the model replies with a structured tool call instead of normal text, such as the tool name plus arguments like
{"city": "Pune"}. It was trained to produce these in the exact required format.The app runs it
The app's code executes the real function: calls a weather API, queries a database, runs code. It can also refuse, or ask you for approval first.
The result goes back
The output is added to the conversation as a tool result. The model reads it and either answers or asks for another tool. Repeating this many times is how an agent works.
- Research roots
- Meta's Toolformer paper (February 2023) showed a model teaching itself to use a calculator, search engines, a translator and a calendar
- Mainstream launch
- OpenAI added function calling to its API in June 2023, with a weather example much like ours
- Common tools today
- Web search, code execution, file reading, image generation, connectors to email, calendars and work apps
- Standard plug
- MCP lets one tool connect to many AI apps in the same way
Tool results are also a way in for attacks: text inside a web page or email can carry hidden instructions. See prompt injection.
Questions people ask
What is function calling in AI?
Function calling is the developer name for tool use. The app describes some functions to the model, the model replies with which one to run and with what inputs, the app runs it and returns the result, and the model uses that to answer.
Does the AI run the tool itself?
No. The model only writes a request in a set format. The surrounding app decides whether to run it, runs it, and passes back the output. This is why apps can ask you to approve actions like sending an email.
What is the difference between tool use and an AI agent?
Tool use is a single ability: asking for one outside action and using the result. An agent chains many tool calls in a loop, deciding each next step itself, to finish a larger goal.
Were ChatGPT plugins the same as tool use?
They were an early consumer version of the idea. OpenAI announced plugins in March 2023 as "tools designed specifically for language models", letting ChatGPT call outside services. Plugins have since been retired, but the same request-and-result tool calling powers today's connectors.