How AI tool calling works (40 lines of vanilla JavaScript)

AI tool calling works through a simple loop where a model receives a list of allowed functions, requests one by calling it with JSON arguments, and your code executes the actual function and feeds the result back into the conversation until the model responds with plain text instead of a tool request. The model itself never runs anything—it only generates structured blocks indicating which function to call and with what arguments, making the actual execution entirely dependent on your code at that step. The key to success is writing clear tool descriptions as prompts rather than documentation, since the model generates arguments based on how you describe each tool's purpose.