How the lab works
Every character in the lab is an AI, not a person. Each one is a digital character built by LucidDataMind and powered by Claude. When you talk to one, this is the pipeline:
- Your browser's speech recognition turns your voice into text.
- The text goes to a server route, and Claude streams the reply back: Claude Haiku 4.5, Claude Haiku 5.5 and Claude Sonnet 5.5, depending on the character.
- Deepgram synthesizes the character's voice on the server, with a Kokoro voice in your browser as the fallback.
- Web Audio effects shape the voice.
- A 3D head rendered with Three.js speaks the words.
Every character's voice is synthesized on the server by Deepgram when available, with Kokoro in your browser as the fallback: only the character's reply text is sent to Deepgram, and the audio streams straight back and isn't stored. Glitch syncs its mouth and its stuttered captions to the audio it is about to play.
Glitch's and Lucid's faces are illustrated portraits of their creator, each turned into a real-time face mesh with MediaPipe — no video, no photoreal cloning. Lucid's is projected as a hologram.
Margin, Brush and Aster are original AI characters; their faces aren't based on any real person.
Your voice is transcribed by your browser to send text to the character; audio isn't stored. Messages are kept for up to 24 hours to run the conversation, then deleted. A shared episode of The Unscripted Room is kept for 90 days, then deleted.
Best in desktop Chrome or Edge. Some mobile browsers fall back to reduced voice mode.