Agora

Agora’s Conversational AI Engine ships LiveAvatar as a first-class avatar provider. You keep Agora’s voice orchestration (ASR, LLM, TTS) and Agora SD-RTN transport — LiveAvatar adds the avatar video participant inside the same channel.

This page assumes you already have an Agora project with Conversational AI enabled and can start a voice agent against an Agora RTC channel. If not, start with Agora’s Conversational AI quickstart, or pick a different integration path.

Upstream reference: Agora docs — LiveAvatar by HeyGen.

Prerequisites

A LiveAvatar account and API key. Sign up at app.liveavatar.com.
An Agora project with Conversational AI Engine enabled, plus an App ID and Basic Auth credentials for the REST API.
A way to mint Agora RTC tokens for at least three participants: the end user, the Agora agent, and the LiveAvatar avatar. The agent and avatar tokens need publish permission; the user token needs both publish (microphone) and subscribe.

Installation

There is no SDK to install on your side. The integration is server-driven: you call Agora’s join REST endpoint with an avatar block, and Agora’s backend brings LiveAvatar into your channel.

Configuration

LiveAvatar is selected by adding an avatar block under properties in the Agora join request body:

{
  "avatar": {
    "vendor": "liveavatar",
    "enable": true,
    "params": {
      "api_key": "<liveavatar_api_key>",
      "avatar_id": "<liveavatar_avatar_id>",
      "quality": "high",
      "agora_uid": "<avatar_rtc_uid>",
      "agora_token": "<avatar_rtc_token>"
    }
  }
}

Parameter	Required	Notes
`api_key`	yes	Your LiveAvatar key from app.liveavatar.com.
`avatar_id`	yes	The avatar to render.
`quality`	yes	`high` (720p), `medium` (480p), or `low` (360p).
`agora_uid`	yes	RTC UID the avatar publishes under. Must be distinct from your `agent_rtc_uid` and any `remote_rtc_uids`.
`agora_token`	yes	RTC token minted for `agora_uid` on the same channel as the agent, with publish permission.
`activity_idle_timeout`	no	Seconds of silence before the avatar tears down. Default `120`.
`disable_idle_timeout`	no	Default `false`. Set `true` to keep the avatar alive indefinitely; overrides `activity_idle_timeout`.

TTS must be 24 kHz. LiveAvatar consumes audio at a 24,000 Hz sample rate. Any other rate in the tts block will error at session start.

How Agora + LiveAvatar fit together

An Agora RTC channel hosts the conversation between three participants:

Participant	Publishes	Subscribes to
End user	Microphone audio	Avatar audio + video
Agora agent (`agent_rtc_uid`)	(no media; forwards TTS audio to LiveAvatar)	User microphone
LiveAvatar (`agora_uid`)	Lip-synced audio + video	TTS audio forwarded by Agora

The avatar is its own RTC publisher. Agora’s Conversational AI Engine routes synthesized TTS audio to LiveAvatar, which lip-syncs and publishes audio + video back into the channel under agora_uid. Your frontend subscribes to that UID for the avatar’s tracks.

How the session starts up

You mint tokens. Issue Agora RTC tokens for the agent UID and the avatar UID against the channel your end user will join.
You POST /join. Call https://api.agora.io/api/conversational-ai-agent/v2/projects/:appid/join with properties.avatar set to the LiveAvatar block above, alongside your llm, asr, and tts configuration. The response returns an agent_id — store it; you will need it to stop the session.
Agora starts the agent. The Conversational AI Engine joins the channel as agent_rtc_uid and starts its ASR → LLM → TTS loop.
Agora spawns LiveAvatar. Using your api_key, avatar_id, agora_uid, and agora_token, Agora’s backend opens a LiveAvatar session and the avatar joins the channel as a publisher.
Your frontend subscribes. The end user joins the channel, subscribes to the avatar’s audio + video tracks, and the conversation runs.

Stopping the session

POST to https://api.agora.io/api/conversational-ai-agent/v2/projects/:appid/agents/:agent_id/leave with the same Authorization: Basic header used for /join, where :agent_id is the value returned from /join. Agora tears down the agent and the LiveAvatar session together.

Minimal example

ElevenLabs TTS at 24 kHz, pointed at LiveAvatar:

curl --request POST \
  --url "https://api.agora.io/api/conversational-ai-agent/v2/projects/$AGORA_APP_ID/join" \
  --header "Authorization: Basic $AGORA_BASIC_AUTH" \
  --header "Content-Type: application/json" \
  --data '{
    "name": "liveavatar-session-1",
    "properties": {
      "channel": "demo-room",
      "token": "<agent_rtc_token>",
      "agent_rtc_uid": "1000",
      "remote_rtc_uids": ["2000"],
      "idle_timeout": 120,
      "asr": { "language": "en-US" },
      "llm": {
        "url": "https://api.openai.com/v1/chat/completions",
        "api_key": "<openai_key>",
        "system_messages": [
          { "role": "system", "content": "You are a helpful avatar." }
        ]
      },
      "tts": {
        "vendor": "elevenlabs",
        "params": {
          "base_url": "wss://api.elevenlabs.io/v1",
          "key": "<elevenlabs_key>",
          "model_id": "eleven_flash_v2_5",
          "voice_id": "pNInz6obpgDQGcFmaJgB",
          "sample_rate": 24000
        }
      },
      "avatar": {
        "vendor": "liveavatar",
        "enable": true,
        "params": {
          "api_key": "<liveavatar_api_key>",
          "avatar_id": "<liveavatar_avatar_id>",
          "quality": "high",
          "agora_uid": "3000",
          "agora_token": "<avatar_rtc_token>"
        }
      }
    }
  }'

agent_rtc_uid, the user UID in remote_rtc_uids, and avatar.params.agora_uid must all be different. Mint a separate token for each.

What changes versus a voice-only Agora agent

Concern	Voice-only Agora agent	With LiveAvatar
Channel participants	User + agent	User + agent + avatar
Where TTS audio surfaces	Published by the agent	Forwarded to LiveAvatar; avatar publishes audio + video
Tokens to mint	Agent + user	Agent + user + avatar
TTS sample rate	Vendor default	Locked to 24 kHz
Frontend rendering	Audio-only subscribe	Must also subscribe to the avatar UID’s video track
Billing	Agora minutes + your inference	Agora minutes + your inference + LiveAvatar minutes

Concurrency and limits

Concurrent sessions are capped by the lower of your Agora Conversational AI limit and your LiveAvatar plan limit. Provision both sides before scaling load tests.

Getting Started

Core Concepts

FULL Mode

LITE Mode

Socials

FAQ

Prerequisites

Installation

Configuration

How Agora + LiveAvatar fit together

How the session starts up

Stopping the session

Minimal example

What changes versus a voice-only Agora agent

Concurrency and limits

Resources

Agora — LiveAvatar by HeyGen

Agora Conversational AI Engine

​Prerequisites

​Installation

​Configuration

​How Agora + LiveAvatar fit together

​How the session starts up

​Stopping the session

​Minimal example

​What changes versus a voice-only Agora agent

​Concurrency and limits

​Resources

Agora — LiveAvatar by HeyGen

Agora Conversational AI Engine

Prerequisites

Installation

Configuration

How Agora + LiveAvatar fit together

How the session starts up

Stopping the session

Minimal example

What changes versus a voice-only Agora agent

Concurrency and limits

Resources