Your website domains
Where your widget is allowed to run. One per line, or comma separated.
Leave empty and it runs anywhere — filling this in means no other
site can load your widget or speak to your account.
example.com also covers
www.example.com and works over
http or https. Use *.example.com
for subdomains. Add a port only if you need to pin one.
Email every conversation when it ends
The whole transcript, who the visitor was, and who answered: the AI, your
team, or nobody. Sent about 15 minutes after the last message. Chats nobody
answered are marked so you can get back to the visitor.
Send to : this website's address, comma-separated. Leave empty to use its admins.
Show the avatar
Off by default. The agent still speaks every reply aloud over the
photo — only the moving face is hidden. Leave this off unless the GPU is running:
it bills by the second, and none of the avatar backends below will connect while
this is unchecked.
Live photoreal avatar (optional, needs a GPU)
The real thing: a GPU box lip-syncs real footage of the person to every sentence she actually says. Everything in the stack is free and self-hosted — the GPU is the only cost. Leave empty and nothing changes.
When it's reachable it takes priority over everything below; when it isn't, the chat silently falls back to the looping clips.
How to set this up
Rent a GPU (an RTX 4090 or L40S is plenty), run bash scripts/runpod_setup.sh in avatar-service/, then build the avatar once with python scripts/build_avatar_pack.py ./sources.
Paste the pod URL above and pick an avatar. Until the GPU is running, the chat keeps using the looping clips below.
Full walkthrough, costs and licence notes: docs/avatar-photoreal-gpu.md and docs/avatar-pack.md.
Video avatar (optional, free & photoreal, fallback)
Two short looping clips of a real person. The chat plays the idle clip and crossfades to the talking one whenever the agent speaks — photoreal, with no GPU, no per-minute cost and nothing to run out of.
Used instead of the flat photo whenever the paid video avatar isn't available, and it takes priority over the 3D model below.
Idle loop — quiet, looking at camera
Talking loop — optional; without it she simply doesn't change while speaking
How to make the two clips
Generate them from the agent photo (free) — SadTalker on Google Colab's free GPU turns one photo into both clips in ~10 minutes: blinking idle face + moving lips. Step-by-step guide: docs/avatar-photoreal-free.md in the repo (start with python scripts/make_avatar_audio.py).
Or record them — 10–20 seconds each on a phone, same seat, same framing, same lighting. Idle: look at the lens, blink, breathe. Talking: read anything at a natural pace. Start and end in the same pose so the loop doesn't jump.
Or generate via a trial — one HeyGen/D-ID free credit renders a talking clip you loop forever.
Or use stock — Pexels and Pixabay have free talking-head footage.
Mute them and keep them small (a few MB). Drop the files in content/agent/ and reference them as /agent-assets/<file>.mp4, or paste any public URL.
Demo mode
What the AI shows when a visitor asks for a demo. In Slides mode it displays the specific slide image that matches the question; in Video mode it plays your demo video; in Live walkthrough mode it opens a page of your site in a window and a cursor clicks and types through it like a real user. Manage your slides & videos on the Demos page.
All — let the AI choose slides, video or walkthrough
Slides — show the matching slide image
Video — play the demo video
Live walkthrough — open the real page and fill it in with a simulated cursor
Branding
Font
The chat's typeface. Served by us, so your visitors' browsers never contact a font service.
Default (Inter)
Roboto
Open Sans
Lato
Montserrat
Poppins
Source Sans 3
Nunito
Work Sans
DM Sans
Raleway
Merriweather
Known pages (AI can redirect here)
Pages on your site the AI may send visitors to. When a visitor asks about a topic that matches one (e.g. pricing), the AI answers, then—after it finishes speaking—takes them to that page. Use a path on the site where the widget is embedded, e.g. /pricing (not a full https:// link).
Add a page
Per-page playbooks
Override the greeting and how fast the chat opens, based on the visitor's page or campaign. The first matching rule wins; pages with no match use the defaults above.
Add a playbook