Skip to main content
Avaluma gives you two ways to run a photorealistic conversational avatar. The fastest is to create an Agent in the Avaluma dashboard at avaluma.ai: you configure the avatar, voice, and behavior in your browser, and Avaluma hosts everything for you. If you need full control over compute and data, you can instead run the pipeline yourself — see Self-Hosting.

Agents (hosted)

Create an agent in the dashboard, share a link, or embed it. Avaluma runs the avatar server and the voice pipeline for you.

Self-Hosting

Run the Avatar Server and LiveKit Agent on your own hardware for full control over data, models, and compute.

What is an Agent?

An Agent is a configured, hosted avatar that lives in your Avaluma account. Each agent bundles everything a live conversation needs:
  • A photorealistic avatar rendered from an Avaluma .hvia model
  • A voice for text-to-speech
  • An AI persona — the system prompt and behavior that shape its responses
  • Sharing settings — a public link and optional embedding on your own website
Once configured, an agent has a stable Agent ID (a UUID) that you use to share or embed it.

Create an agent

1

Sign in to the dashboard

Log in at avaluma.ai and open the Agents section of your dashboard.
2

Create a new agent

Click New agent and give it a name. As soon as it has a name, the agent is created. From then on the editor saves every change automatically. Then select the avatar you want it to use. Your available avatars come from the .hvia files issued with your license.
3

Configure voice and persona

Choose a voice for text-to-speech and write the agent’s instructions (system prompt). The instructions define the agent’s tone, knowledge, and behavior during a conversation. You can write them into the form yourself or let the prompting guide walk you through them. A default voice is already preselected — a voice must be selected for the agent to speak.
4

Test

Click Test agent at the bottom and start a test conversation to confirm the avatar speaks and responds as expected.

The agent editor

All settings of an agent sit on one page, one below the other: Avatar, Voice, Prompting, Background knowledge, Deployment, Metadata and, at the very bottom, the collapsible Advanced area.

Automatic saving

There is no save button. The editor saves every change on its own, about one second after your last input and immediately when you leave a field. The bar at the bottom of the page shows the state: Saving … or Saved.
  • Every change takes effect from the next conversation, for embedded agents too.
  • The editor does not save incomplete input, for example a tool without a name. The bar then shows Not saved: with the reason. Clicking it jumps to the field. Everything else is saved anyway.
  • If saving fails, for example without an internet connection, the error appears in the bar. Try again saves once more.

Prompting guide

The Prompting section holds the form for the instructions with the fields Greeting, Identity, Goal, Output Formatting, Guardrails, User Information and Additional Information. You can fill it in yourself. Above it, Start prompting guide starts a conversation with an avatar that walks you through the instructions and fills in the form for you. The browser needs access to your microphone for this.
  • The fields are locked during the conversation. You can watch live what the guide enters: a changed field lights up briefly.
  • Stop prompting guide ends the conversation. After that you can edit all fields yourself again.
  • Uploading, importing and deleting in Background knowledge is locked in the meantime.

Advanced

The Advanced area is collapsed by default. It holds the Avatar ID, the models for speech recognition (STT) and replies (LLM), Turn taking and Tools. The defaults fit most agents.

Configure an agent

Open an agent to adjust its configuration at any time. Common settings include:
The direct link (/agent/AGENT_ID) and the kiosk are reachable for every agent, so treat the Agent ID like a public link. Only embedding via iframe needs to be enabled separately — together with the origins it is allowed on.

Choosing a speech model

Under Voice you can also pick the ElevenLabs model that generates the speech. Every model offered works with your own instant voice clones (voice ID).

Voice settings

Under Voice Settings you adjust Stability, Similarity and Speed with sliders. Each slider has a Voice default checkbox above it. While it is checked, the value is not sent and ElevenLabs uses the voice’s own setting. Uncheck it to enable the slider; it then starts at the ElevenLabs defaults: stability 0.5, similarity 0.75, speed 1.0. Which sliders appear depends on the model:
The sliders adapt automatically once you switch the model: unsupported ones disappear, and on v3 a stability value in between is rounded to the nearest preset.
New agents start on v4 Turbo. Existing agents keep their model until you change it yourself.

Defaults for new agents

A newly created agent starts with this configuration. You change the speech output under Voice and everything else in the Advanced area: If you switch speech recognition away from Deepgram Flux, turn-taking is reset to “Not configured”, because STT endpointing only works with Flux.

Manage agents

In the Agents overview, each agent has a menu (three dots) with additional actions:
Duplicating also carries over the embedding settings (including the allowed origins). The copy is therefore immediately reachable under the same conditions as the original.

Per-agent usage statistics

In the Agents overview, each agent card shows how many conversation minutes that agent has been used in the current month. The statistics icon on the card opens a window with the key metrics: A trend additionally shows conversations over time, switchable between 1 day (by hour), 7 days, and 30 days. Use the date field to pick the end date of the period (default: today) and look back at past days or weeks.
The monthly total across all agents and your plan limit are shown in the bar at the top of the Agents page — see Pricing & Billing for details.

Share and embed

Each agent is reachable at a hosted URL based on its Agent ID:
For troubleshooting, the direct link understands the same diagnostic parameters as embedding and the kiosk: audiodebug, webaudio and speakerphone. To put the agent directly on your own site, embed it as an <iframe> and control it from your page.

Embedding

Embed an agent on your website with an <iframe> and customize its appearance and behavior via query parameters.

Messaging

Control the embedded avatar programmatically and receive events with the postMessage API.

Kiosk mode

Run the agent fullscreen on an unattended display — with a screen wake lock and idle auto-reset.