No server · no API keys · runs on your GPU

In-Browser Model Lab

Six small apps and games where the AI runs on your own laptop, inside the browser tab. Nothing you type or say leaves your machine.

Checking WebGPU… Checking model cache… Checking microphone support…
Game

Prompt Golf

Make a tiny language model hit a target output in the fewest characters. Nine holes with banned words, hints that cost strokes, and a course editor.

~120–790 MBno mic
Experiment

Nervous Narrator

Ask a question and the model answers out loud. Its voice slows, quiets and wobbles on the words it's least sure of, using its own token probabilities.

~880 MB – 1.9 GBmic optional
Tool

Slides That Follow You

Present without a clicker. It listens, matches what you say to each slide with embeddings, and advances when you get there. Talk notes afterwards.

~105 MBmic
Game

AI Telephone

Pass a sentence down a chain of AI relays that speak, mishear over a crackly line, and retell it. A drift chart shows the meaning falling apart.

~980 MBmic optionalsound on
Game

AI Debate Club

Two small models with ridiculous personas argue a motion out loud. A third model judges, with probabilities, and the room votes on an applause meter.

~210 MB – 2 GBmic optionalsound on
Game

Outage at 3 A.M.

A text adventure for the on-call crowd. The engine owns the truth; a tiny embedding model understands what you type, a small LLM narrates, and a grounding check flags what it invents.

0 MB to play+23 MB to type anything+790 MB narratormic optional
Next

What should we build?

Voice-controlled diagrams, talking to your data, Jargon Bingo, a page tone-shifter and more, with notes on how each would work.

Before you start