Letting the model reach for a tool
Function calling, also called tool use, is the mechanism by which a model recognizes that its own knowledge is insufficient and invokes an external capability — search, calculation, a database query. It compensates for the weaknesses of LLMs with instruments, and it sits at the center of how AI is used today.
How it runs
Asked what to wear for today’s weather, the model does not know the weather. So it decides to call a weather tool, invokes it with a location and date, reads the returned data, and recommends clothing based on it.
To the user it just looks like a good answer. Underneath, the model chose an instrument, used it, and folded the result into its response.
What it enables
- Answering from current information via web search, which also reduces hallucination
- Accurate calculation and data analysis through a calculator or code execution
- Real actions: creating calendar events, sending mail, updating a database
- Chaining several tools to complete a piece of work — the basis of AI agents
How MCP fits
MCP standardizes how tools are called across services. If function calling is the model’s ability to use instruments, MCP is the common socket that any instrument can plug into. Chat assistants that search, generate images, or run code are all doing this, which explains the mechanics behind announcements that an assistant can now do something new.
How well a model chooses the right tool at the right moment has become a measure of capability in its own right. When a benchmark lists tool use or agentic performance, this is what it means.
As a user, a note in an answer saying it searched or ran a calculation tells you a tool call happened.