🐾 Golden Bowl

Somewhere in this maze is a golden bowl of food. Three dogs and a pig have to find it — dogs have a great nose but poor eyesight, the pig sees far but can barely smell. Each animal can be played by you, a simple Bot, or an LLM agent (any provider) calling look, move and recall tools in real time.

🐶
🐶
🐶
🐷
Same number = same maze. Replaying one maze with the models swapped around is the fair way to compare them.
How the setup works

Each LLM animal has its own connection, so different animals can run different models or providers. The three dogs share the same senses, so putting the models you're comparing on the dogs makes the race fair; the pig sees farther but smells less. Only one animal can be steered by arrow keys at a time — Tab / 1-4 switch, M mutes. LLM agents call each provider directly from your browser, so that provider's normal usage charges apply. Providers must allow browser (CORS) requests; for one that doesn't (DeepSeek is reported not to), run python3 cors-proxy.py and pick the "via local relay" preset. Protocol, URL, model and thinking level are remembered per animal; API keys never are.