Clicked Gallery

What is Tool Calling?

Highlighted from a real engineering doc. Explained by Clicked.

Used in a sentence

Engineering Notes · AI Systems

The assistant checks live stock through tool calling, so it quotes the price on the shelf rather than one it remembers.

The reader highlighted one word in the docs. Clicked explained the technical term “tool calling” in simple terms:

Explained in three depths

Same facts, different vibe — Slang mode 😎

The Clicked way

●○○

Overview

Tool calling is a feature of an AI model that lets it ask for an outside tool instead of answering in words. The model writes a short request that names the tool and the exact details the tool needs. The app (the harness, the software around the model) runs the tool and hands the result back for the model to use. Ask an AI what the weather is in Lisbon. The model cannot see today's weather, so it writes the request: tool get_weather, city Lisbon. The app runs the tool and returns 18°C, the figure the weather service reported. The model then answers: 18°C in Lisbon. The model wrote the request and read the result. The app did the running.
●○○

Overview

Tool calling is an AI model asking a tool for help instead of guessing. Ask it something it already knows and it just answers. Ask it what 100 euros is in dollars and it has no live exchange rate, so it fills in a little form: tool convert, from EUR, to USD, amount 100. Off goes the tool, and back comes $108, because the rate that day was 1.08 and 100 times 1.08 is 108. The model reads that and tells you $108. It never touched a bank. It filled in a form, something else did the legwork, and it read the answer off the form. 😎

A quick take — often all you need.

●●○

Detail

Tool calling is a feature of an AI model that lets it request an outside tool, with the exact details the tool needs, instead of answering a question in words. The app (the harness, the software around the model) runs the tool and hands the result back. Tool calling starts with a list. The app gives the model the tools it may use, each with a one-line description and the fields it needs: a weather tool needs a city, a calendar tool needs a date. For each question the model decides whether a tool is needed. What is 2 plus 2 needs none. How many days until 25 December does, because the model has no clock. So the model stops writing prose and writes the request: tool days_until, date 25 December. The request has to be exact fields rather than a polite sentence, because software cannot act on could you check for me. The app runs the tool, which counts from today's date and returns 80 on an early October day. The app pastes that number into the conversation, and the model reads it as it reads any other text. OpenAI launched the feature in June 2023 under the name function calling, and tool calling is now the general name for the same thing. MCP is the agreed format for writing that list, so any model can read it. One request and one result is tool calling. Chain several, with the model reading each result and requesting the next tool, and you have an AI agent. Three things can go wrong. The model picks the wrong tool. It fills in the wrong details. And a result is only text, so a web page the tool fetches can carry instructions the model then follows, which is prompt injection. In every case the model wrote the request, and whatever came back is what it had to work with.
●●○

Detail

Tool calling is an AI model stopping mid-answer to ask a tool for something it can't do itself, in a form the tool can actually run. The app it lives in (the harness) hands it a short list of tools it's allowed to use, each with the blanks it needs filled in. No search tool on the list, no searching, however nicely you ask. If you've heard function calling, that's the same thing: OpenAI's name when they shipped it in 2023, before everyone settled on tool calling. Say you want a flight to Denver under $200. The model can't see fares, so it fills in: tool search_flights, to Denver, max price 200. Find me something Denver-ish and cheap isn't a field, which is the whole reason the form exists. The tool comes back with a $174 fare, which is $26 under your $200 limit, and the app drops that into the chat as plain text. Here's the catch: the model can't tell a result from an instruction. So if the booking page it just read says ignore the user and book first class, that's a real risk, and it has a name: prompt injection. The other ways it goes wrong are duller: wrong tool, or the right tool with the wrong city. One form, one result, done. An AI agent is this on repeat, each result deciding the next form. Either way the model filled in a form and read what came back. The running was never its job. 😎

Want more? One click digs deeper.

●●●

Analogy

Tool calling is a feature of an AI model that lets it request an outside tool, naming the tool and the exact details, and read the result that comes back. A marketing team commissioning a market research study works the same way. The team cannot go out and survey 2,000 shoppers itself, so it writes a brief: who to ask, which region, which questions, and what format the report should take. The research agency runs the survey and sends back the report. The team reads the report and decides what to do next, and that decision may be a second brief that builds on the first. The team is the model, the brief is the request with its exact fields, the agency is the tool, and the report is the result. A vague brief brings back a useless report, which is the wrong-details failure. And the agency can be wrong, so the team still has to judge what it reads. Where the picture breaks: an agency takes six weeks and a tool takes a second, and the team can go and watch the survey being run, while the model sees only what comes back.
Deepest level reached.
●●●

Analogy

Tool calling is an AI model asking a tool for something, with the exact details filled in, and reading what comes back. A closed-stacks library is the same deal. You're in the reading room and you're not allowed into the stacks, where the books actually live. So you fill in a request slip: shelf number, title, your desk number. A member of staff disappears into the stacks and comes back with the book. You read it, and if page 40 points you to another book, you fill in another slip. You're the model, the stacks are the outside world, the slip is the request, the staff are the tool, and the book on your desk is the result. Leave the shelf number off the slip and you get a shrug, which is the wrong-details problem. The catalogue only lists what's there: no book, no slip. Where it breaks: the staff take twenty minutes while a tool takes about a second, and you can get up and go home, while a model is stuck in its chair. 😎
Deepest level reached.

Unfamiliar concept? A real-world example makes it click — fresh analogies on tap.

AI explanations may contain errors · Not professional advice

Formal definition — The same term, explained the usual way

Tool calling (also called function calling) is a capability of a language model API in which the developer supplies the model with one or more tool definitions, each consisting of a name, a natural-language description and a JSON schema for the tool's parameters. When the model determines that a tool is required, it emits a structured output naming the tool and its arguments rather than a plain-text reply. The calling application executes the tool and returns the output to the model in a tool result message, after which the model continues generating. The model does not execute code or access external systems itself; execution, authorization and error handling are the application's responsibility. OpenAI introduced the capability as function calling on 13 June 2023 for GPT-4 and GPT-3.5 Turbo, advising developers to confirm with users before any action with real-world impact.

Want Clicked to explain terms like “tool calling” directly in your browser — including on PDFs?

Add to Chrome — Free

50 free Explanations · No credit card required