Function calling

A structured-output capability of large language models in which the model returns a JSON-shaped tool-call rather than free-form text — naming a function, supplying validated parameters, and signalling that the calling application should execute the function and return the result. The mechanism that turns a model into an agent.

Function calling is the OpenAI-native term for what Anthropic calls 'tool use' and Google calls 'function declarations'. The protocols differ; the operating concept is the same. The Model Context Protocol (MCP) is the emerging open standard that abstracts vendor-specific function-calling formats into a portable surface. Procurement teams evaluating vendor lock-in should ask whether the deployment's tool-use surface is vendor-portable (MCP) or vendor-specific (proprietary function-calling format) — that single answer determines the cost of switching foundation models later.

Articles that analyse this term

• MCP and the coming standard for enterprise agent tooling
• AI assistant vs AI agent: the procurement distinction

Primary sources

• OpenAI. Function calling (API documentation)
• Anthropic. Tool use with Claude