Tool Calling
Tool Calling means that a language model doesn't just write text, but specifically calls external programs — such as a calculator, a search engine, or calendar software. The model itself decides when it needs such a tool, and incorporates the result into its answer.
A language model like ChatGPT can really only do one thing: generate text, word by word. It doesn’t know what time it currently is, and it can’t look anything up on the internet. With Tool Calling, it therefore gets access to external programs that take on exactly such tasks. The model then doesn’t write an answer for the user, but rather a kind of order slip: “Call the weather program, location Berlin.” A surrounding system carries out this request and hands the result back. Only after that does the model formulate its actual answer. Such helper programs are called tools — hence the name.
Why a language model alone isn’t enough
A language model has two built-in weaknesses. First, its knowledge is old. It comes from texts that were collected before training. It doesn’t know about yesterday’s events. Second, it doesn’t really calculate, but guesses the most probable continuation of a sentence. With a multiplication of large numbers, this often goes wrong.
Tool Calling fixes both at once. For current information, the model calls a search engine. For calculations, a calculator or a small piece of program code. The result is then not guessed, but measured or computed. The difference is significant: a model without tools can present a wrong number completely convincingly.
Even more important is the second step. A tool can not only read, but also act. It can send an email, enter an appointment, or trigger an order. This turns a chatbot into a system that actually gets tasks done. This is precisely what the so-called AI agents, which have been widely reported on since 2024, are based on.
From order slip to finished answer
The process has fixed stages. At the start, the model receives a list of the available tools. Each is briefly described: name, purpose, and what information it needs. A weather tool, for instance, requires a place name and a date. This list sits invisibly ahead of the user’s actual question.
When a question now comes in, the model decides: answer itself, or use a tool. If it needs help, it doesn’t output a normal sentence, but a strictly formatted data record with tool name and parameters. This format is important, because a program cannot understand free text. The surrounding system — not the model itself — executes the call.
The result is then presented to the model like a new message. Now it formulates the answer for the human. This often repeats itself several times: first search, then calculate, then summarize. A common misconception is the idea that the model itself runs the programs. It only suggests them. This separation is a safety measure, because it allows an operator to check or block every call.
Weather apps, travel booking, and the MCP standard
In everyday life, you encounter Tool Calling constantly without it being named. When a chatbot tells you the current stock price, it has queried a price database. When it generates an image, it has called an image model as a tool. Voice assistants on phones also work this way when they set an alarm.
In companies, the use is greater. An assistant in customer service looks up the order status in the company database and initiates a refund. So that such connections don’t have to be rebuilt every time, there has been an open standard for this since 2024: the Model Context Protocol, or MCP for short. It describes uniformly how a model and a tool talk to each other.
In news reports, the term mostly appears in connection with agents. When a company announces that its AI can now “book an entire trip,” behind this there is almost always a chain of tool calls. And when reports cover security vulnerabilities in AI, they are often precisely about this interface. A model that is tricked by cleverly worded text into making unwanted calls is a real risk.