Any camera works here — a webcam, phone-as-webcam, or a virtual camera. Only needed if you pick Camera as the input source.
The agent can speak short first-person lines out loud (GPT-Live). Use an earphone if you have one — with speakers, the mic is muted while the agent talks so it doesn't hear itself.
GPT-Live voice plays through the selected output. No selection means the system default device.
Screen share: share the window with your coding problem or design doc — not this page, or the agent will watch its own Pad. Camera: point it at yourself or a whiteboard and just talk through what you're working on.
Auto detect is the default. A manual pick only hints new code generation — clear on-screen evidence always wins.
One session at a time. If someone else is running the demo you'll see a notice — share the viewer link below and they watch live from any network.