Large language models excel at text but can't interact with external systems. Tool calling bridges this gap by letting the model request actions from your application. The model decides what to do, but your code executes the tool, maintaining a clean separation between reasoning and execution. This architecture prevents debugging prompts instead of system behavior.

In practice, tool calling operates as an iterative loop. The model receives a user query, checks defined tools, and returns a structured Tool Call object. Your application runs the tool and sends results back, potentially triggering another tool call in a chain. This structured protocol replaces guesswork, as the model must specify exact tools and parameters.

Implementations like NodeLLM formalize this with typed tool classes. The model uses metadata—name, description, and schema—to decide which tool to invoke. Vague descriptions lead to hallucinations or ignored tools. Production systems must also handle parallel calls, failures, and timeouts, which live in your runtime, not the model.