Anva avatars for LiveKit Agents

Give your LiveKit voice agent a face. The agent keeps its own STT, LLM and TTS; Anva joins the room as a second participant and speaks the agent's words with the lips in time, video and audio published together.

pip install livekit-plugins-anva

from livekit import agents, rtc
from livekit.agents import Agent, AgentSession
from livekit.plugins import anva, deepgram, openai, silero

async def entrypoint(ctx: agents.JobContext):
    session = AgentSession(
        stt=deepgram.STT(), llm=openai.LLM(), tts=openai.TTS(), vad=silero.VAD.load(),
    )
    avatar = anva.AvatarSession(avatar_id="av_...")   # or preset_id="..."
    await avatar.start(session, room=ctx.room)
    await session.start(agent=Agent(instructions="You are a friendly guide."), room=ctx.room)

Set ANVA_API_KEY (from anva.ai → API keys) and the usual LIVEKIT_URL, LIVEKIT_API_KEY and LIVEKIT_API_SECRET. ANVA_AVATAR_ID can stand in for avatar_id.

What happens

1. start() mints a room token for the avatar participant (kind agent, attribute lk.publish_on_behalf naming your agent) and asks Anva to create an avatar_only session that joins your room with it. Anva answers once the avatar is in the room with its video and audio tracks published.
2. The agent's audio output is replaced by LiveKit's DataStreamAudioOutput, so speech goes to the avatar over the room's data stream rather than straight to the room. Interruptions and playback receipts use LiveKit's standard lk.clear_buffer, lk.playback_started and lk.playback_finished RPCs.
3. aclose() (called for you when the job shuts down) ends the Anva session.

Billing is Anva's avatar_only rate per connected minute. The audio can be any sample rate and channel count; Anva converts it to what its lip-sync takes.

Options

A 400 from Anva (wrong mode, bad avatar, plan limits) raises AnvaException without retrying; a 503 no_capacity is retried after Anva's Retry-After.

Front end

Your front end sees two agent participants. LiveKit's components pick the avatar's tracks automatically; by hand, it is the participant whose lk.publish_on_behalf attribute equals your agent's identity.

Tests

livekit-server --dev &
LIVEKIT_TEST_URL=ws://127.0.0.1:7880 python -m pytest tests