Realtime TTS-2 is live. Built for realtime conversation that feels human. Read the Realtime TTS-2 announcement

Realtime API overview

Build low-latency voice agents over WebSocket or WebRTC, with turn detection and interruptions handled for you.

The Realtime API connects STT, the LLM you choose, and TTS in one spoken conversation. It uses one WebSocket or WebRTC connection and handles turn detection and interruptions. It follows the OpenAI Realtime protocol, so you can migrate an OpenAI Realtime client with few changes. Inworld extensions add back-channel responses, fillers, and memory.

WebSocket quickstart

A Node.js server and a browser client, streaming audio both ways.

WebRTC quickstart

Voice using the browser's built-in support, with automatic echo cancellation.

Realtime Playground

Talk to a voice agent in the Portal.

JS examples

Complete Node.js and browser projects.

Python examples

Complete Python projects.

API reference

Every event, for WebSocket and WebRTC.

Using AI to code? Give your assistant the docs index at https://docs.inworld.ai/llms.txt. For live search, add the MCP server.

Prefer the terminal? Install the Inworld CLI with npm install -g @inworld/cli. Use it to create and manage API keys for the Realtime API. AI agents can use it too.

Your first session

Connect and set up a session, then stream audio or text and handle the responses:

text
wss://api.inworld.ai/api/v1/realtime/session?key=<session-id>&protocol=realtime
  1. Connect with Authorization: Basic <api-key> from a server, or a session token from a browser. You receive session.created.
  2. Send session.update with your instructions, model, voice, and tools.
  3. Stream audio with input_audio_buffer.append, or text with conversation.item.create.
  4. Handle response.output_* events until response.done.

Explore

Configuring models

Choose the STT model, the LLM or router, the voice, and turn detection.

Tool calling

Let the agent call your functions mid-conversation.

Managing conversations

Conversation items, turn detection events, interruptions, and errors.

Adding naturalness

Steering, back-channel responses, and fillers that keep the agent feeling present.

Twilio integration

Put the agent on a phone number.

For voice output without speech input, use Voice responses. Generate and speak an answer in one HTTP request.