Experiment: Real-time voice conversation with GPT-Live

[sktkkoo]

I got GPT-Live running in Yorishiro through Codex app-server.

This is still an early experiment and is not included in the current release. I built this prototype to explore what it would feel like to have real-time conversations through a Yorishiro resident, rather than interacting only through text in the terminal.

What works in the prototype

• Microphone audio is streamed to GPT-Live over WebRTC
• GPT-Live responds with voice
• The active Yorishiro avatar lip-syncs to the spoken response
• GPT-Live connects to the same Codex app-server thread as the terminal
• Existing Yorishiro reactions and self-referential MCP tools continue to work during the conversation

gpt-live-test.mp4

gpt-live-color.mp4

Current limitations

The current integration with Yorishiro is still partial. The regular Codex agent receives the existing Yorishiro persona, but that persona is not yet passed directly to the GPT-Live conversational layer.

Feedback

I would love to hear which parts of this experience seem promising and what you would expect from voice interaction with a terminal agent.