Skip to main content
By the end of this guide you will have the Avaluma Avatar Server rendering your .hvia avatar on a GPU, a LiveKit Agent running a full voice AI pipeline (STT → LLM → TTS), and a connected test client where you can speak to your avatar in real time.

​
Prerequisites

Confirm you have the following before you begin:
  • Docker & Docker Compose installed on your server
  • NVIDIA GPU (CUDA 12, min 6 GB VRAM) with the NVIDIA Container Toolkit installed
  • An Avaluma license key and at least one .hvia avatar file
  • A LiveKit account (Cloud or self-hosted) with an API key and secret
1

Deploy the Avatar Server

Clone or download the avatar-server example directory, then place your .hvia avatar files where the container can find them:
mkdir -p avatar-server/assets/avatars
cp /path/to/your-avatar.hvia avatar-server/assets/avatars/
Open avatar-server/docker-compose.yaml and set a secure password for the utility API:
environment:
  - API_SERVER_HOST=api.avaluma.ai   # public IP or domain of this server
  - API_UTILS_PWD=CHANGE_THIS        # replace with a strong password
Start the server:
docker compose up -d
The Avatar Server is now available at http://localhost:8080. For production, add the optional Caddy reverse proxy included in avatar-server/reverse_proxy/ to terminate TLS automatically.
The container uses pull_policy: always, so restarting it will automatically pull the latest Avaluma image.
2

Configure the LiveKit Agent

Clone or download the livekit-agent example directory. Copy the environment template:
cp .env.example .env.local
Open .env.local and fill in your credentials:
AVALUMA_LICENSE_KEY=""      # Your Avaluma license key
AVATAR_SERVER_URL="https://your-avatar-server.com"
# Avaluma Hosted: https://api.avaluma.ai

LIVEKIT_URL=""              # wss://your-livekit-instance.livekit.cloud
LIVEKIT_API_KEY=""          # API key from your LiveKit project
LIVEKIT_API_SECRET=""       # API secret from your LiveKit project
If you are using Avaluma’s managed Avatar Server, set AVATAR_SERVER_URL=https://api.avaluma.ai. If you deployed your own server in Step 1, use its public URL.
Next, open agents/1-agent-with-livekit-inference/agent-1.py and set avatar_id to the filename of your .hvia file without the .hvia extension:
avatar_id = "your-avatar-id"   # e.g. "studio-avatar-v2" for studio-avatar-v2.hvia
The AvatarSession block in that file wires the voice pipeline to the Avatar Server:
from avaluma_livekit_plugin import AvatarSession

avatar = AvatarSession(
    license_key=license_key,       # Your License Key
    avatar_id=avatar_id,           # Avatar identifier (Name of .hvia file)
    avatar_server_url=avatar_server_url,
)
await avatar.start(agent_session=session, room=ctx.room)
You do not need to modify this block — it reads the values you set in .env.local and avatar_id automatically.
3

Start the Agent

Run only the conversational agent (agent-1):
docker compose up livekit-agent-1 -d
To start both example agents at once, omit the service name:
docker compose up -d
Each agent must have a unique AGENT_NAME in docker-compose.yaml when you deploy multiple agents to the same LiveKit project.
4

Test the Connection

With all three services running — Avatar Server, LiveKit server, and the agent — connect a client to your LiveKit room.
Open the Avaluma Test Client in your browser. It works with any LiveKit setup — self-hosted or cloud. Enter your LiveKit server URL and credentials, then click Connect to start speaking to your avatar.
Your avatar should appear in the video track and respond to your voice in real time.

​
Next Steps

Architecture

Learn how the Avatar Server and LiveKit Agent interact under the hood.

Avatar Server

Configure GPU resources, run multiple avatars simultaneously, and set up HTTPS.

LiveKit Agent

Swap in different STT, LLM, and TTS models and add custom agent logic.

External Audio

Stream audio from any external service directly to the avatar without an AgentSession.