Inspiration

Most AI in VR demos are chat boxes in space. We wanted a desk you can think with — a multilingual AI companion that talks, looks things up on the web, and builds 3D for you in blocks, then hands it over so your hands do the craft. Prototyping a shelf, a stage set, a machine part, or a toy should feel like sketching on a real desk: talk it into existence, then sculpt.

What it does (Productivity)

AgentAsia Voxel is a hands-first, seated spatial prototyping studio for Meta Quest 3.

Talk to Naga in your language. Voice in, voice out. English, Japanese, French, and more. Domo arigato and Naga replies in kind. Captions included. STT + TTS on device.
Live knowledge (Tavily + LLM). Ask anything: weather, news, sports, prices, how-tos. Naga answers from the live web — not a frozen training set. Build commands and factual Q&A share one conversation.
Build almost anything you ask, in blocks. Ten blueprints plus open-ended LLM builds. Layer-by-layer assembly on your real desk in passthrough. Then your hands finish it: pinch the roof as one cluster, slide, delete, undo. Agent drafts; you sculpt.
3D modeling at multiple levels of detail. Draft 12x12 up to Mosaic 64x64 — smaller blocks for engineering mockups and artistic models. Export the desk as standalone HTML to share or review in a browser.
Nine desk games when you need a break: Tetris, Sling, Star Swarm, Marble Maze, Whack-a-Gloom, Breakout 3D, Block Tower, Flap Bro, and Mike Penguin. Same desk, hands-first, short sessions.

How we built it

WebXR + Three.js in the browser (no APK), seated and FoV-aware. Hands-first pinch / grab / snap-to-grid / cluster grab / undo / trash — no controllers required. AgentAsia voice stack: AssemblyAI STT, Nemotron LLM, Tavily web search, Kokoro TTS. Voxel engine (InstancedMesh), 10 blueprints, grid presets, HTML export. Live: https://naga-atelier.vercel.app

Challenges

Making hand tracking feel precise (snap + audio cues), keeping 60 fps with voxels plus games, teaching pinch without walls of text, and shipping ten loops on one desk without feeling like a demo dump. Agent placement is a draft — human hands always win.