How avatarx works

An agent is the unit you build and deploy. It carries a face, a voice, a personality and a prompt, optionally a knowledge base of your own content, and a set of tools it may use mid-conversation. Once it is published you drop one script tag on any page and visitors can talk to it.

1. 1

   Create an agent

   Go to Agents and press New agent, or start from one of the templates (Lila, Alex, Emma, Jack) to pre-fill the wizard.
2. 2

   Work through the five tabs

   Avatar, Voice, Behavior, Knowledge and Conversation. Only a name and a prompt are required; everything else has a sensible default and stays editable later.
3. 3

   Try it in the preview

   The right-hand panel talks to the real model with the prompt you are writing, so you can feel the tone before you create anything.
4. 4

   Create, then copy the embed snippet

   Creating the agent indexes any knowledge you staged and hands you the script tag.
5. 5

   Test it in the sandbox

   Launch sandbox on the agent card runs the real widget against your agent, so you see exactly what a visitor will get.

The wizard and the agent settings page hold the same fields, so anything you set while creating can be changed afterwards.

Avatar

Name, a short description, and the face. The gallery lists the built-in video avatars; the one you pick streams live and lip-syncs to the agent's speech.

Voice

• Language — the language the agent speaks.
• Voice — pick one and press preview to hear it.
• Pronunciation dictionary — for names, brands and acronyms it gets wrong. Give the word and how to say it (AVAT → ay-vat).

Behavior

• Agent role and personality — a sentence of context and a tone.
• Agent prompt — the instructions. Write only the instructions here: role, personality, topics to avoid and the response-length cap are folded in automatically at call time. Use View the composed prompt on the settings page to see the exact text the live agent receives.
• Creativity — from more predictable to more diverse. It sets the model temperature; the settings page shows the resulting value.
• Model — GPT-4o, GPT-4o mini, Claude 3.5 Sonnet or Haiku.

Conversation

Give the agent your own content and it answers from that instead of guessing. Everything you add is chunked, embedded and searchable — and stays editable.

Sources

• Files — PDF, TXT, MD or DOCX, up to 20 MB each.
• Pasted text — FAQs, policies, price lists. Saved as its own document so you can edit it later.
• A website — index a single page, or follow links to crawl up to 50 pages, 2 levels deep. Re-crawl any time.

Strict or hybrid

Strict means the agent answers only from the knowledge base and politely declines anything it does not cover — the right setting for support and compliance. Hybrid lets it interpret around the facts conversationally.

Editing what it knows

Open Knowledge, pick an agent, and you can edit any document's text, rename it, re-index it, or delete it. Saving re-chunks and re-embeds that document and removes the old vectors, so a fact you delete stops being retrievable rather than lingering in the index.

Tools are real functions, not prompt suggestions. A disabled capability is not merely discouraged — the model has no way to call it.

Always on

Capabilities you switch on per agent

Scheduling rules

When appointments are on you set the timezone, appointment length, gap between bookings, earliest bookable time, how far ahead visitors may book, and business hours — including a second window per day for a lunch break. The preview under the form shows the exact slots visitors will be offered, which is the fastest way to catch hours that are too tight.

Connect Google Calendar and every booking gets its own calendar event and a unique Meet link. Without it, bookings reuse the static Meet room you set on the agent.

Custom webhook tools

Under Tools → Custom toolsyou can define an HTTPS endpoint the model may call: give it a snake_case name, a description (this is what the model reads to decide when to call it), the URL and method, an optional JSON Schema for its arguments, and credentials if it needs them. Then attach it to an agent from that agent's Tools tab.

Agents can show things, not just talk about them. Add a flyer, business card, menu or price list on the agent's Media tab and it appears as a card in the chat panel when it is relevant.

• Images, PDFs, MP4/WEBM or an external link — up to 25 MB each, 50 per agent.
• The title and trigger keywords are how the agent decides an item is the right one, so name things the way visitors ask for them.
• Show at the start puts an item on screen right after the greeting — good for a business card.

Copy the snippet from the agent's Embed tab and paste it just before the closing </body> tag.

<!-- avatarx embeddable widget -->
<script src="https://avatarx.net/components/widget/widget.js" async></script>
<script>
  window.VoiceAgentConfig = {
    agentId: "YOUR_AGENT_ID",
    apiUrl: "https://avatarx.net"
  };
</script>

Options

What the visitor gets

• A launcher in the bottom-right that expands into the avatar.
• Mic and speaker toggles, a chat tab, and an end-call button.
• A chat panel with live transcripts, your conversation starters, and media cards.
• A clean end screen when the agent closes the call — plus a star rating when you enable the feedback screen.

Reading the feedback rating

When the feedback screen is on, the widget dispatches a DOM event on the host page. Nothing is stored server-side yet, so capture it yourself if you want to keep it:

window.addEventListener("voiceagent:feedback", (e) => {
  // { agentId, score: 1..5 | null, comment: string | null }
  console.log(e.detail);
});

• Conversations — every session with its transcript.
• CRM — leads the agent captured and appointments it booked, both editable, with CSV export for leads.
• Insights — volume, duration and engagement over time.

Outbound email

Appointment confirmations and handoff alerts are sent from your own SMTP server, so they come from your domain. Add the host, port and credentials under Settings and send a test — until that test passes, handoffs are recorded but nobody is emailed.

Google Calendar

Connect once for the whole account. Every booking then gets its own calendar event and Meet link.

Security

Sign-in creates a server-side session you can revoke. Under Settings you can see active sessions, sign out other devices, and change your password — which revokes every other session.

Live conversations consume credit minutes from your account balance. The sidebar shows what is left. When the balance runs out the widget stops connecting new calls and the API answers 402 credits_exhausted, so top up before a launch.

Building against the API instead? The API reference covers authentication and every endpoint.