WebRTC - Inworld AI Documentation

Connect via WebRTC for browser-native, low-latency voice. A WebRTC proxy bridges your peer connection to the same realtime service used by the WebSocket transport, transcoding OPUS ↔ PCM16 and forwarding events transparently.

Endpoint

https://api.inworld.ai

Endpoint	Method	Description
`/v1/realtime/calls`	POST	SDP offer/answer exchange
`/v1/realtime/ice-servers`	GET	STUN/TURN server configuration

Authentication

Pass your Inworld API key as a Bearer token. The proxy forwards it to the realtime service.

Authorization: Bearer <base64-api-key>

Keep the API key server-side. Serve it to the browser via a backend endpoint (see examples below).

Flow

Fetch config from your server (API key + ICE servers)
Create RTCPeerConnection with ICE servers
Create data channel oai-events + add microphone track
Create SDP offer → POST it with optional initial session config to /v1/realtime/calls → set SDP answer
Data channel opens → send any later session.update patches → start conversation

Audio flows via RTP tracks (no manual encode/decode). Events flow via data channel using the same JSON schema as WebSocket.

Session config

You can configure the session during signaling or through the data channel:

During signaling: include the session object in a JSON or multipart /v1/realtime/calls request. The proxy sends it upstream as an initial session.update before returning the SDP answer. It requires a successful send but does not wait for session.updated.
After connection: send session.update through the data channel. Partial updates patch the initial configuration.

For example, a JSON signaling request can opt into OpenAI-style tool continuation:

const response = await fetch(`${PROXY}/v1/realtime/calls`, {
  method: 'POST',
  headers: {
    Authorization: `Bearer ${API_KEY}`,
    'Content-Type': 'application/json'
  },
  body: JSON.stringify({
    sdp: offer.sdp,
    session: {
      type: 'realtime',
      providerData: {
        auto_tool_response: false
      }
    }
  })
});

const { sdp } = await response.json();
await pc.setRemoteDescription({ type: 'answer', sdp });

You can also send the same session fields later as a WebSocket-compatible session.update through the data channel. See Configuring Models for the full breakdown of STT, LLM, and TTS configuration.

dc.send(JSON.stringify({
  type: 'session.update',
  session: {
    type: 'realtime',
    model: 'openai/gpt-4o-mini',
    instructions: 'You are a concise concierge.',
    output_modalities: ['audio', 'text'],
    audio: {
      input: {
        turn_detection: {
          type: 'semantic_vad',
          eagerness: 'medium',
          create_response: true,
          interrupt_response: true
        }
      },
      output: {
        voice: 'Clive',
        model: 'inworld-tts-2',
        speed: 1.0
      }
    },
    tools: [{
      type: 'function',
      name: 'get_weather',
      description: 'Fetch weather for a location',
      parameters: {
        type: 'object',
        properties: { location: { type: 'string' } },
        required: ['location']
      }
    }],
    providerData: {
      stt: {
        voice_profile: true,
        language_hints: ['en-US', 'es-MX'],
        end_of_turn_confidence_threshold: 0.7,
        min_end_of_turn_silence: 200,
        max_turn_silence: 5000,
        vad_threshold: 0.5
      },
      tts: {
        segmenter_strategy: 'sentence',
        steering_handling: 'emit_once',
        language: 'en-US',
        delivery_mode: 'CREATIVE',
        conversational: false
      },
      memory: {
        enabled: true,
        turn_interval: 5,
        max_facts: 50
      },
      backchannel: {
        enabled: true
      },
      responsiveness: {
        enabled: true
      }
    }
  }
}));

providerData carries Inworld-specific extensions to the OpenAI-compatible session shape — STT tuning, TTS segmentation/steering, automatic memory, and more. Most providerData fields are hot-swappable via partial session.update, but a few are read only at session open and ignored afterwards (notably providerData.tts.conversational and providerData.tts.user_turn_mode). See Inworld Realtime API Extensions for the field-by-field reference.

Audio

Unlike WebSocket (manual base64 PCM), WebRTC handles audio natively:

Input: browser captures mic and sends OPUS over RTP automatically
Output: proxy sends AI audio back as an RTP track — attach to <audio> to play

pc.ontrack = (e) => {
  const audio = document.createElement('audio');
  audio.autoplay = true;
  audio.srcObject = new MediaStream([e.track]);
  document.body.appendChild(audio);
};

response.output_audio.delta events are not sent through the data channel — audio is delivered via the RTP track instead.

Text & Responses

Same as WebSocket, but sent through the data channel:

dc.send(JSON.stringify({
  type: 'conversation.item.create',
  item: { type: 'message', role: 'user', content: [{ type: 'input_text', text: 'Hello!' }] }
}));
dc.send(JSON.stringify({ type: 'response.create' }));

Events

Same event types as WebSocket, received on the data channel.

Option 1: Direct WebRTC

Server — serves the page and a /api/config endpoint that fetches ICE servers and keeps the API key hidden:

import 'dotenv/config';
import { readFileSync } from 'fs';
import { createServer } from 'http';

const html = readFileSync('index.html');
const API_KEY = process.env.INWORLD_API_KEY || '';
const PROXY = 'https://api.inworld.ai';

const server = createServer(async (req, res) => {
  if (req.url === '/api/config') {
    let ice = [];
    try {
      const r = await fetch(`${PROXY}/v1/realtime/ice-servers`, {
        headers: { Authorization: `Bearer ${API_KEY}` },
      });
      if (r.ok) ice = (await r.json()).ice_servers || [];
    } catch {}
    res.writeHead(200, { 'Content-Type': 'application/json' });
    res.end(JSON.stringify({ api_key: API_KEY, ice_servers: ice, url: `${PROXY}/v1/realtime/calls` }));
    return;
  }
  res.writeHead(200, { 'Content-Type': 'text/html' });
  res.end(html);
});
let port = 3000;
server.on('error', (e) => {
  if (e.code === 'EADDRINUSE') { console.warn(`Port ${port} in use, trying ${++port}…`); server.listen(port); }
  else throw e;
});
server.listen(port, () => console.log(`http://localhost:${port}`));

Client — full WebRTC flow in the browser:

const cfg = await (await fetch('/api/config')).json();
const stream = await navigator.mediaDevices.getUserMedia({ audio: true });

const pc = new RTCPeerConnection({ iceServers: cfg.ice_servers });
const dc = pc.createDataChannel('oai-events', { ordered: true });
stream.getAudioTracks().forEach(t => pc.addTrack(t, stream));

pc.ontrack = (e) => {
  const audio = document.createElement('audio');
  audio.autoplay = true;
  audio.srcObject = new MediaStream([e.track]);
  document.body.appendChild(audio);
};

dc.onopen = () => {
  dc.send(JSON.stringify({
    type: 'session.update',
    session: {
      type: 'realtime',
      model: 'openai/gpt-4o-mini',
      instructions: 'You are a helpful voice assistant.',
      output_modalities: ['audio', 'text'],
      audio: {
        input: { turn_detection: { type: 'semantic_vad', eagerness: 'medium', create_response: true, interrupt_response: true } },
        output: { voice: 'Clive', model: 'inworld-tts-2' }
      }
    }
  }));
};

dc.onmessage = (e) => {
  const msg = JSON.parse(e.data);
  if (msg.type === 'response.output_text.delta') console.log(msg.delta);
  if (msg.type === 'error') console.error(msg.error?.message);
};

const offer = await pc.createOffer();
await pc.setLocalDescription(offer);
// wait for ICE gathering...
const res = await fetch(cfg.url, {
  method: 'POST',
  headers: { 'Content-Type': 'application/sdp', Authorization: `Bearer ${cfg.api_key}` },
  body: pc.localDescription.sdp,
});
await pc.setRemoteDescription({ type: 'answer', sdp: await res.text() });

Option 2: OpenAI Agents SDK

The OpenAI Agents SDK manages the full WebRTC lifecycle — peer connection, SDP exchange, mic, and audio playback:

import { RealtimeSession, RealtimeAgent, OpenAIRealtimeWebRTC } from '@openai/agents/realtime';

const agent = new RealtimeAgent({
  name: 'assistant',
  instructions: 'You are a helpful voice assistant.',
  model: 'openai/gpt-4o-mini',
});

const cfg = await (await fetch('/api/config')).json();
const audioEl = document.createElement('audio');
audioEl.autoplay = true;

const session = new RealtimeSession(agent, {
  transport: new OpenAIRealtimeWebRTC({
    useInsecureApiKey: true,
    audioElement: audioEl,
    changePeerConnection: async (pc) => {
      if (cfg.ice_servers?.length) pc.setConfiguration({ iceServers: cfg.ice_servers });
      return pc;
    },
  }),
  model: 'gpt-4o-realtime-preview-2025-06-03',
});

await session.connect({ url: cfg.url, apiKey: cfg.api_key });
session.sendMessage('Hello!');

The server-side /api/config endpoint is identical to Option 1.

WebSocket vs WebRTC

	WebSocket	WebRTC
Audio	PCM16 base64 (manual)	OPUS via RTP (native)
Latency	Higher	Lower (UDP)
NAT	Not needed	ICE (STUN/TURN)
Events	WS messages	DataChannel (same schema)
Best for	Server-side / Node.js	Browser voice apps

API reference for full event schemas.

​Endpoint

​Authentication

​Flow

​Session config

​Audio

​Text & Responses

​Events

​Option 1: Direct WebRTC

​Option 2: OpenAI Agents SDK

​WebSocket vs WebRTC

Endpoint

Authentication

Flow

Session config

Audio

Text & Responses

Events

Option 1: Direct WebRTC

Option 2: OpenAI Agents SDK

WebSocket vs WebRTC