An API (Application Programming Interface) is a standardized calling interface agreed upon between software components: the caller passes parameters in a predetermined format, and the callee returns results, with neither party needing to understand the other's internal implementation. It forms the foundation of modern software division of labor and serves as the core channel for large models to connect with the external world, supporting operations such as queries, transactions, communication, and data read/write.

These specifications enable tool capabilities to be parsed by machines, thereby supporting model discovery and combination of tools. For models, APIs grant them execution capabilities. In 2023, Patil et al. in the paper Gorilla: Large Language Model Connected with Massive APIs fine-tuned the LLaMA model and integrated a document retriever, enabling the model to generate calls based on the latest documentation. On APIBench, which covers HuggingFace, TorchHub, and TensorHub, their call accuracy surpassed GPT-4 and significantly reduced hallucinations.

APIs extend models from text generation to actual operations on external systems, forming the basis for tool calling and agent implementation. The resulting engineering constraints include least privilege, call behavior auditing, and interception mechanisms for potentially dangerous operations.