A large language model, on its own, can only read text and write text back. Tool use, sometimes called function calling, is what breaks that limit: the model is given a list of available actions, such as “look up an order,” “send an email,” or “query this database,” and it can choose to call one, pass it the right arguments, and read back a real result before deciding what to say or do next.
This is the mechanism that turns a chatbot into something that can touch a business system. It is also where the practical risk of an AI project concentrates, because a model that can only talk can only be wrong in a sentence, while a model with tool use can take a wrong action in a real system. Most agentic AI incidents trace back to a tool being available that should have required a human check first, not to the underlying model being insufficiently capable.
The Model Context Protocol standardises how a model discovers and calls tools across different vendors and systems; tool use is the underlying capability that protocol is built to support. Judge a proposed AI agent by its exact list of callable tools, not by how capable the model behind it sounds in a demo.