Describe the model in plain words. A coding agent writes the recipe —
named parts, real dimensions — antics-modelkit checks it and
exports a .glb that Godot, Unity, Unreal and the browser load natively.
Free, and the model comes back as code you can keep changing.
Paste this to Claude, Cursor or any coding agent that can run Node — then swap the watchtower for whatever your game needs. The link teaches it the kit.
Read https://antics.gg/modelkit/llms.txt, then build me a low-poly wooden watchtower with a ladder and a lamp that can swing, and export it as a .glbOne function, the kit's own verbs, every part named. You never have to read it — but you can, and so can the agent, which is the whole trick: it can look at what it wrote and change one number instead of starting over.
export function watchtower(h = 2.4) {
return { category: "prop", parts: [
{ name: "legs", geometry: legs(h), color: 0x8a6b45 },
{ name: "cabin", geometry: cabin(h), color: 0xa3805a },
{ name: "lamp", geometry: lamp(h + .3), color: 0xf2c14e },
] };
}Inside-out surfaces, holes you can see into, parts floating off the body, a silhouette that turns to mush at play size — the build catches what a glance misses, and what comes out opens in any engine like any other asset.
See what comes out →Hand the agent a drawing or a reference photo with the same prompt and it models what it sees. It reconstructs by reasoning, not by scanning — our pocket watch was built from one photograph, movement and all.
The watch, from a photo →Everything above assumes you bring an agent. We're building the hosted version — type the
prompt on this site, we run the agent, you download the .glb and the recipe.
It is not live yet, and its page doesn't pretend otherwise.
Text-to-3D here means flat-coloured, low-poly-leaning game assets with named parts. It will not scan your photo into that exact object, and it won't hand you a photoreal dragon — for textured hero pieces, a mesh generator is honestly the better tool.
When to use which →The kit is MIT, the recipes are MIT, the catalogue models carry CC licences, and there's no account or credits. The one thing you bring is your own coding agent — its tokens are between you and it.
A first pass in a couple of minutes. The point is what happens after: previews render in half a second, so you look, say “wider” or “less lamp” in plain words, and look again until it's right.
You say what's wrong, in words. Because the model is code, the fix is an edit to a number or a part — not a fresh generation that loses everything that was already right. The invisible failures, the kit's checks catch before you ever see them.
Any coding agent that can run Node: Claude Code, Cursor, Codex and friends. Point it at antics.gg/modelkit/llms.txt and it learns the kit on the spot.
The models are half of it. Vibe code the game around them and antics wires in rooms, live sync and a leaderboard, then hosts it at a link your friends can open. Free to start.