Function Tools

Function tools let the model ask your app for data or actions during a conversation. Your app runs the function and returns the result. The model then speaks the answer.

Define tools

Add tools to session. Each tool uses the OpenAI GA shape.

{
  "type": "session.update",
  "session": {
    "type": "realtime",
    "tools": [
      {
        "type": "function",
        "name": "get_order_status",
        "description": "Look up the status of a customer's order by order ID",
        "parameters": {
          "type": "object",
          "properties": {
            "order_id": { "type": "string", "description": "Order ID, for example ORD-1042" }
          },
          "required": ["order_id"]
        }
      }
    ],
    "tool_choice": "auto"
  }
}

`tool_choice`	Behavior
`auto` (default)	The model decides whether to call a tool
`none`	The model never calls a tool
`required`	The model must call a tool on this turn

Handle a call

The loop has four steps:

Collect the arguments

The model streams arguments as response.function_call_arguments.delta. Wait for response.function_call_arguments.done, which has the full JSON string and call_id.

{
  "type": "response.function_call_arguments.done",
  "call_id": "call_abc123",
  "name": "get_order_status",
  "arguments": "{\"order_id\":\"ORD-1042\"}"
}

Run your function

Parse arguments and run your code. Keep this on your server or in a trusted backend.

Return the output

Send the result as a function_call_output item that uses the same call_id:

{
  "type": "conversation.item.create",
  "item": {
    "type": "function_call_output",
    "call_id": "call_abc123",
    "output": "{\"status\":\"shipped\",\"eta\":\"2026-10-24\"}"
  }
}

Ask for the reply

Send response.create. The model speaks the answer using the output.

{ "type": "response.create" }

Example event flow

client  → session.update            (tools defined)
client  → input_audio_buffer.append (user asks "Where is order 1042?")
server  → response.function_call_arguments.done
client  → conversation.item.create  (function_call_output, same call_id)
client  → response.create
server  → response.output_audio.delta … response.done

Notes

output must be a string. Return JSON as a string.
Return an error message in output when your function fails, so the model can apologize.
Keep tool results short. Long results make the reply slower.

MCP tools inside a Realtime session are not supported yet. Define function tools that call your own backend.

Client events for conversation.item.create and response.create
Server events for function call events

Get Started

Guides

Reference

Function Tools

Define tools

Handle a call

Example event flow

Notes

​Define tools

​Handle a call

​Example event flow

​Notes

​Related

Define tools

Handle a call

Example event flow

Notes

Related