What is function calling?

Function calling is how a model gets from words to actions. Without it, connecting a model to a system meant asking for text and parsing it with fragile code. With it, the model fills in a typed request, such as a customer ID and a date range, and your code decides what to do with it. Almost every AI agent is built on this mechanism.

It also marks the safety boundary. The model does not execute anything. OpenAI’s guide has the application execute code using the tool call input, and Anthropic’s documentation says your application executes client tools. That gap between the model asking and the code acting is where you check permissions, validate arguments and ask a person.

You start by describing each function: a name, a description of when to use it, and a JSON schema for its parameters. You send these definitions with the request. The model reads the user’s message and the descriptions, and either answers directly or returns a tool call naming a function and its arguments. In Anthropic’s API the response arrives with a stop reason of tool_use and a tool_use block.

Your code runs the function, then sends the result back to the model in a second request. The model uses it to answer, or asks for another call. OpenAI describes this as five steps, from the first request with tools to the final response. Both vendors let you steer the choice: let the model decide, require a call or force a specific function, and OpenAI also lets you restrict calls to an allowed subset. Both support several tool calls in one turn, and both let you turn that off.

A schema is a request to the model unless you switch on a strict option. OpenAI’s strict mode makes function calls adhere to the schema instead of on a best-effort basis, with rules such as setting additionalProperties to false. Anthropic offers a strict option for custom tools. Validate arguments in your own code either way.

You define one function, get_order_status, described as “Look up the current status of a customer order by order number”, with one required string parameter, order_number. A customer writes: “Where is my order AB-2231?”

The model returns a call to get_order_status with order_number set to AB-2231. Your code checks that the caller owns the order, queries the order system and sends back that the parcel left the warehouse yesterday. The model replies to the customer in plain language. Had the customer left out the order number, a clear description and a strict schema make it more likely the model asks for it rather than inventing one.

Swfte Connect handles tool calling on its OpenAI-compatible chat completions endpoint, so code that already sends OpenAI-style tool definitions can point at Connect and keep the same shape. Whether a given model supports tool calls, and how well, depends on the model and provider, so check before you route tool-heavy traffic to it.

In Swfte Studio you rarely write tool definitions by hand. Integrations give an agent its tools, and an approval step can hold a run for a named person before a consequential tool call. Studio agents can also call tools on external MCP servers.

• Swfte Connect: OpenAI-compatible API in front of many providers
• Swfte Studio
• Tools for AI agents

Definitions on this page were read on the sources below on 7 October 2026. Where sources define the term differently, the page says so. The full glossary lists more terms.