What is Function Calling?

A raw language model can only produce text — it can describe what it thinks should happen, but it can't itself check a calendar, query a database, or send an email. Function calling closes that gap: the application defines a set of available functions with names, descriptions, and typed parameters, the model is given that list alongside the user's request, and when appropriate the model responds not with prose but with a structured call — "invoke get_weather with location=Boston" — which the application executes and feeds the result back into the conversation. This is the mechanical foundation underneath most AI agents and the Model Context Protocol (MCP): an agent that can "do things" rather than just talk is, under the hood, a language model repeatedly deciding which function to call, executing it via the host application, and reasoning over the result to decide the next step. Reliability depends heavily on how well the function schema is described — vague names or missing parameter descriptions lead to malformed or wrong calls — and on the host application validating and sandboxing what a called function is actually allowed to do, since the model's judgment about when to call something is not itself a security boundary.