The point is that you call the LLM to generate the code that lets you talk to the API, rather than writing that glue code yourself. Not that you call the LLM to talk to that API every time.