01
Media path
OpenAI WebRTC audio
Encrypted microphone input and synthesized agent speech flow as WebRTC media.
Xirsys × OpenAI Realtime
Use OpenAI Voice Agents on Xirsys
Talk to an OpenAI voice agent with real-time WebRTC audio. Xirsys supplies the STUN and TURN infrastructure that keeps the session reachable across all networks.
WebRTC
Live audio
STUN/TURN
NAT traversal
Data channel
Agent events
Voice session
Ready to connect
idle
Required · BYOK
Sent once over HTTPS to mint an ephemeral Realtime token, then discarded. It is not saved or logged. Use a temporary or restricted project key and revoke it after testing.
Connection events
- WebRTC route
- Not connected
- Selected path
- No candidate pair selected
- Xirsys TURN
- Not in use
- Xirsys WebSocket
- Disabled
One realtime session
Each transport has one clear job.
The browser connects directly to OpenAI for media and agent events. Xirsys handles the network path, providing a relay only when a direct connection cannot make it through.
02
Network path
Xirsys NAT traversal
Short-lived STUN/TURN credentials supply host, reflexive, and relay candidates.
03
Control path
Realtime data channel
OpenAI JSON events travel over the oai-events WebRTC data channel.
Open source SDK
Clone it. Configure it. Ship voice.
The complete TypeScript server SDK, browser WebRTC client, demo, tests, and implementation guide are available under the MIT license.
git clone https://github.com/xirsys/openai-realtime-xirsys-webrtc.git
Browse the repository →