Keyframe Labs Plugin

[kradkfl]

Keyframe Labs API Docs

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no potential bugs to report.

View in Devin Review to see 5 additional findings.

[tinalenguyen]

thank you for the PR! could you also add the plugin to this pyproject file as well: https://github.com/livekit/agents/blob/main/livekit-agents/pyproject.toml

[tinalenguyen]

just wondering, are there no plans to support setting the emotion via an API call? i think that would be more ideal if possible

[kradkfl]

Do you mean by way of a REST call to api.keyframelabs.com?

If so, the latency would be pretty high. Additionally, the avatar session itself on the backend isn't listening to any central server calls once connected to the room, it's just listening to data channels (which is what we use here).

Are you imagining that the user of the avatar plugin wouldn't want to access the underlying avatar object and call this function themselves? Or that accessing the underlying avatar isn't ergonomic?

[tinalenguyen]

ah yes, i was just curious how users would update settings mid-session in other scenarios/use cases, but the latency using the data channels is much better as you said. iirc most of our other providers don't allow for dynamic updates like this, so this would be a new feature (very cool to see the changes in real-time!)

[tinalenguyen]

looks good to me! small ask - could you update the examples to use this setup instead for the AgentSession:

session = AgentSession(
        stt=inference.STT("deepgram/nova-3"),
        llm=inference.LLM("google/gemini-2.5-flash"),
        tts=inference.TTS("cartesia/sonic-3"),
        resume_false_interruption=False,
    )

i noticed that this setup called set_emotion more often as well, so it really seems like the avatar is reacting throughout the conversation. i didn't notice much of a difference in latency either