What is tool calling?
Tool calling is when a language model asks a program to carry out a function, such as searching a system, fetching a file or sending a message.
A language model can only write text. Given descriptions of tools, it can write a request for a particular tool to be used, with the details needed. The program around the model makes the call and hands back the result.
Tool calling is what turns an AI assistant into an AI agent. The model reads the result, chooses the next step and calls another tool if needed, until the task is done.
The tools can be anything from a search in a knowledge base to a change in a business system. With a protocol such as MCP, the same tool can be used by several AI applications.
How it works
Here is an example. Someone asks when a contract expires. The model sees that it has a tool for searching the contract register, writes a request with the contract's name and gets the date back. The answer to the user is based on what the tool returned.
The model never makes the call itself. The application decides whether the call is allowed, with which permissions and whether someone needs to approve it first.
Why it matters
Every tool widens what the model can affect. A wrong answer can be corrected, whereas a wrong call may already have moved money, deleted files or sent data to the wrong recipient. That is why controls do most good at the calls themselves.
What to keep in mind
Log every call with what was requested, on whose behalf and what happened. An audit log makes it possible to follow an agent's steps afterwards.
Give tools least privilege. A tool that reads the calendar does not need to send email, and protocols such as MCP make it easier to keep tools apart.
What it means for you
Tell apart tools that read from tools that change things. Those that change something, or send something outside the organisation, should require approval.
Common questions
Is tool calling the same as an AI agent?
Tool calling is the building block. An AI agent uses it over several steps to complete a task. Read more about AI agents.
Who decides which tools the model may use?
Whoever builds or configures the AI application. The model can choose only from the tools the application offers.
Sources: Tools in the MCP specification
Related terms
Read more
All terms Updated