A minimal voice agent powered end-to-end by LiveKit Inference. The playground exposes STT, LLM, and TTS pickers that swap the corresponding component live via RPC, so you can hear how each provider feels in the same conversation without restarting the session.

How the live swap works

The playground sends an RPC on every dropdown change:

room.localParticipant.performRpc({
  destinationIdentity: agent.identity,
  method: "set_stt_model" | "set_llm_model" | "set_tts_model",
  payload: JSON.stringify({ value: "deepgram/nova-3" }),
});

The agent registers one handler per control. All three (STT, LLM, TTS) call update_options(model=...) to swap the active model without restarting the session.

Run locally

pip install -r requirements.txt
python agent.py dev

The model list shown in the playground is sourced from examples/playground.yaml. To add or remove options, edit the controls block on the inference example there.