Wander
Apps & Workflows · Side Project · 2026
How much of an idea gets lost on the way to a text box? Wander steers a live world model with voice, gesture, and expression instead.
A prompt bar makes creating single-player. One person types, everyone else watches. Wander takes the interface off the keyboard and puts it in the room. Voice, expression, and a pointed hand feed a live world model at once, so the scene answers to whoever is talking. World models already read the world. This one reads the makers.
Three senses
Three channels run at once. The Web Speech API streams words as they are spoken, so a scene can turn mid-sentence. face-api.js reads the face seven ways and reports how sure it is. MediaPipe Hands tracks 21 landmarks, which is what turns a pointed finger into a place on the screen.
Weighting the senses
The three channels never carry equal weight. Voice sets the story, a gesture interrupts it, a face only tints what is there. Confidence picks the wording, so a hesitant smile writes cheerful and a certain one writes ecstatic. Pointing crops the frame at the fingertip and runs it through a vision model, which is how "make this blue" arrives naming the object.
Four hours
Wander was built in four hours at the 2026 Odyssey Hackathon, with Yves Chen on the Odyssey integration and Renzo Marsino on the story. The demo ran live for Odyssey's CEO and Soleio, both judging. No settings, no onboarding, nothing to learn. Walk up, talk, point. It didn't place.
The time limit did the editing.
Odyssey has since retired the API, so Wander ships as a repo you run yourself. Clone it, add your own key, and the three senses drive whatever world model you point them at. [github.com/brunomarsino/Wander](https://www.github.com/brunomarsino/Wander).