How AI Actually Calls an API? Tool Calling Explained from Scratch

This article provides a foundational guide to understanding 'tool calling' in AI models. Contrary to the misconception that models execute code independently, the author explains that models act as decision-makers, generating structured requests that applications then execute. The process follows a four-step loop: sending a prompt with tool descriptions, receiving a structured request, executing the function, and returning the result for a final answer. The author demonstrates this using Amazon Bedrock and Python, contrasting tool calling with simple context injection for static data like dates. Furthermore, the piece introduces the Model Context Protocol (MCP) as a standardized way to manage and discover tools, moving away from manual hardcoding. By mastering this loop, developers can transition from static models to dynamic agents capable of interacting with real-world APIs, weather services, and databases, effectively bridging the gap between frozen training data and live, actionable information.