An A2UI renderer whose output device is a speaker.
One payload drives a screen and a voice at the same time, and saying
“the second one” emits the identical action a tap would have — so the agent
cannot tell which happened, and doesn’t need to.
Live · needs an API key
A real Claude agent composing surfaces at runtime. The catalog is compiled into the model’s output schema, so it cannot emit a component you didn’t approve.
Scripted · no key
The four moments on rails: mid-turn switch, screen off, the readback gate, and swapping the speech engine.