Voice Agents

Overview

Voice Agents let you build low-latency spoken interfaces on top of OpenAI speech-to-speech models. The SDK keeps the Realtime API mental model intact, but wraps the raw event flow in RealtimeAgent, RealtimeSession, and transport helpers that make tools, guardrails, handoffs, and session history easier to work with.

Under the hood, the same Realtime concepts from the official Realtime API with WebRTC, Realtime conversations, and voice activity detection guides still apply. The Voice Agents SDK adds a TypeScript-first layer on top of that API so you can stay focused on product logic instead of rebuilding transport and event handling from scratch.

Note

Start with the Quickstart if you are connecting a browser client with WebRTC. Jump to the transport guide first if you are deciding between WebRTC, WebSocket, SIP, Twilio, or Cloudflare Workers.

Start here

Voice Agents Quickstart Build your first realtime voice assistant using the OpenAI Agents SDK in minutes.

Building Voice Agents Learn how session lifecycle, VAD, interruptions, multimodal input, tools, and history work in the SDK.

Realtime Transport Layer Choose between WebRTC, WebSocket, SIP, or a custom transport, and know when to drop down to raw events.

What the SDK adds

• Browser-first WebRTC setup with ephemeral client tokens.
• Server-side WebSocket and SIP transport options.
• Automatic interruption handling and local conversation history updates.
• Multi-agent orchestration through realtime handoffs.
• Function tools, hosted MCP tools, approvals, and delegation patterns.
• Output guardrails and tracing support for live spoken interactions.

Pick the next page

If you need to…	Go here
Connect a browser client safely with WebRTC and ephemeral tokens	Voice Agents Quickstart
Understand session lifecycle, VAD, interruptions, image input, tools, and history	Building Voice Agents
Decide between WebRTC, WebSocket, SIP, and custom transports	Realtime Transport Layer
Run a phone or telephony experience on Twilio	Realtime Agents on Twilio
Connect from Cloudflare Workers or other workerd runtimes	Realtime Agents on Cloudflare

Why speech-to-speech

Speech-to-speech models process user audio directly, so you do not have to build a separate speech-to-text, text reasoning, and text-to-speech chain for every turn. That keeps latency down and makes interruptions, mixed text and voice input, and tool calls feel much more natural in realtime applications.