3D avatar
React Three Fiber, ARKit blendshapes, 56 Mixamo animations. Expressive and performant.
Voice I/O
Web Speech API input, ElevenLabs TTS output. Zero external dependencies for the input path.
18 widget types
Status, code, metric, progress, chart, terminal, table, diff, action, and more — spawned contextually during conversation.
The Conversation Compiler
A small neural rewriter (gpt-4.1-nano or comparable) that sits between the agent and the TTS pipeline. Input: raw agent output (markdown, JSON, technical text) Output: spoken-language version plus emotion, animation, and widget metadataExample
Agent output:Deployed to Railway. Build 47 passed. 3 of 5 tests green.After Conversation Compiler:
- Spoken: “Deploy’s live. Build passed, three of five tests green.”
- Emotion: smile
- Animation: Head-Nod-Yes
- Widgets:
[status: "Deploy" → success],[progress: "Tests" → 60%]
Widget system
18 contextual widget types spawn during conversation:
Widgets appear next to the avatar’s speech, stay for a configurable duration, and can be interactive (approve/deny buttons route back through the Gateway).
Governance integration
Avatars reacts to policy events from the SINT Protocol gateway. Expression and widget choice depend on the tier and outcome.Hardware
Browser
WebGL 2.0, modern browser. Works on laptops, tablets, and phones. The avatar scales to the viewport.
Kiosk / desktop
Electron wrapper for dedicated operator kiosks. Voice wake word, always-on microphone, full-screen avatar mode.
Deployment
.env.
Read next
SINT OS
How Avatars composes with the rest of the stack.
GitHub repo
Source, build instructions, contribution guide.