glove-voice-livekit

LiveKit as an adapter in the Glove voice stack — two halves that share one room connection:

• LiveKitTransport — the room leg every LiveKit-backed voice host otherwise hand-rolls: join, publish the agent's voice as a paced WebRTC track, feed remote mic tracks back out as PCM events, carry JSON on the data channel. Barge-in is server-authoritative: clear() flushes the outbound AudioSource queue, so there is no client playback buffer to chase. attachRealtime(rt, transport) binds it to a glove-voice-s2s RealtimeAgent in one call.

• LiveKit avatars — TavusLiveKitAvatar and AnamLiveKitAvatar implement the glove-voice-avatar AvatarAdapter contract (and pass its conformance suite), so a face over LiveKit is interchangeable with the Daily-based Tavus echo adapter: attachAvatar(rt, avatar) and done. Under the hood they speak LiveKit's published avatar protocol — the provider's worker joins your room as a second participant (token kind agent, lk.publish_on_behalf pointing at your agent) and publishes synchronized voice+face itself; agent PCM reaches it over the lk.audio_stream byte stream, and barge-in is the lk.clear_buffer RPC. A glove agent is indistinguishable from a LiveKit Agents worker as far as the avatar can tell.

Voice only

import { LiveKitTransport, attachRealtime, mintParticipantToken } from "glove-voice-livekit";

const transport = new LiveKitTransport({
  url: process.env.LIVEKIT_URL!,
  token: await mintParticipantToken(
    { apiKey: process.env.LIVEKIT_API_KEY!, apiSecret: process.env.LIVEKIT_API_SECRET! },
    { roomName: "call-42", identity: "agent" },
  ),
});
await transport.connect();
attachRealtime(rt, transport);       // mics → model, model → track, interrupt → flush
await rt.start();

With a face

import {
  TavusLiveKitAvatar, TAVUS_AVATAR_IDENTITY, mintAvatarToken,
} from "glove-voice-livekit";
import { attachAvatar } from "glove-voice-avatar";

// The avatar publishes the voice on the agent's behalf — don't double it.
const transport = new LiveKitTransport({ url, token, publishAgentAudio: false });
await transport.connect();
attachRealtime(rt, transport, { agentAudio: false });

const avatar = new TavusLiveKitAvatar({
  apiKey: process.env.TAVUS_API_KEY!,
  faceId: process.env.TAVUS_FACE_ID!,   // minimal echo PAL ensured automatically
  livekitUrl: url,
  avatarToken: await mintAvatarToken(creds, {
    roomName: "call-42",
    identity: TAVUS_AVATAR_IDENTITY,
    onBehalfOf: "agent",
  }),
  wire: transport.avatarWire(TAVUS_AVATAR_IDENTITY),
});
await attachAvatar(rt, avatar);       // connects + bridges speech/end/interrupt

AnamLiveKitAvatar is the same shape (avatarId instead of faceId, ANAM_AVATAR_IDENTITY); written against Anam's documented API and the conformance suite, live verification pending a key (#71).

Session renewal: providers can end an avatar session under you — Anam's plan cap force-ends conversations at 3/5/10 minutes below Growth tier. The worker leaving the room (participant_disconnected on its identity) is the signal; disconnect() + connect() on the adapter mints a fresh session into the same room. examples/livekit-rooms does this automatically with a debounce.

Browsers see the avatar as an ordinary room participant: attach its video track and you have the face — no provider SDK on the client.

See examples/livekit-rooms for the full layered setup (front agent + mesh worker + station rooms) on this transport, with the avatar as an env toggle.

Status

Wire protocol validated against LiveKit Agents' published plugins and @livekit/rtc-node type declarations; conformance + unit tests green (pnpm test). Live verification tracked on #72.