Also called: function calling, tool use · Last checked: 7 November 2026
In short: Tool calling is the ability of an AI model to ask other software to do something, such as search the web, check stock or send an invoice. The model picks the tool and fills in the details.
Table of Contents
How it works
- Describe the tools. Developers give the model a list of tools, saying what each does and what details it needs.
- The model asks. When a request needs a tool, the model replies with the tool name and the details, such as a customer ID.
- The software acts. Your system runs the tool, returns the result to the model and the model uses it in its reply.
Where you’ll meet it
When ChatGPT searches the web or Gemini checks your calendar, the model is calling a tool. Most AI agents rely on it.
What it means for your business
- AI that acts. Tool calling lets AI work with your own systems, such as your order database or booking calendar.
- You choose the tools. The model can only use the tools you provide, so you decide what it can and cannot do.
- Start read-only. Give tools that look things up before tools that change records or move money.
Watch out for
- Wrong details. A model can pick the right tool but fill in the wrong amount or customer. Validate inputs in your software.
- Too much access. A tool with broad permissions can be misused. Use restricted keys.
- Tricked into acting. Hidden instructions in emails or web pages can push a model to call tools you never intended.
Example
Stripe. On 14 November 2024 the payments company released its agent toolkit, which turns Stripe actions such as creating payment links into tools an AI agent can call. Stripe said it works with any AI model that supports function calling, and advised firms to use restricted keys so agents could only reach the features they needed.
Tool calling vs Model Context Protocol
| Tool calling | Model Context Protocol | |
|---|---|---|
| What it is | A model’s ability to request a tool | A shared standard for connecting tools to AI apps |
| Set up | Defined by each developer for each app | Built once, works with many AI apps |
| How they relate | The basic skill | A common way to offer tools for that skill |
Frequently asked questions
Does the AI model run the tool itself?
No. The model only writes a request naming the tool and the details. Your software decides whether to run it, which is where you add checks and limits.
How many tools can one AI model use?
Dozens in practice, but accuracy drops as the list grows. Give each task a short list of clearly named tools with plain descriptions.
Can tool calling make payments safely?
It can, with safeguards. Limit amounts, use test mode first, require a person to approve payments and keep a log of every call.