An engine LLMs can actually build games on

Why vibe coding 3D games collapses — and what fixes it

Ask an AI assistant for a Three.js game and you'll get something impressive in minutes — one sprawling file of scene setup, game loop, and input handling. Ask for the tenth feature and it starts breaking the first nine. The problem isn't the model; it's the absence of a contract. An engine supplies exactly that: components with defined lifecycles, scenes as validated data, and clear boundaries the model can't wander across. That's the difference between prompt-to-playable and prototype-to-production.

Plain TypeScript components — a contract a model can follow

Behavior in Threenity is an ordinary TypeScript class extending Component, with schema-defined fields and lifecycle hooks. No codegen, no hidden runtime magic: what the model reads is what runs. The component contract — one class, one responsibility, fields declared up front — is the same guardrail that keeps human codebases sane, and models follow it just as well.

Schema-validated JSON scenes — deterministic structure on disk

Scenes and prefabs are versioned JSON, validated with Zod on every load and merge-friendly by design. An agent can read a scene, diff it, and edit it directly on disk — and if it writes something invalid, validation catches it at the boundary instead of corrupting the world. Deterministic structure in, or nothing in.

Embedded docs: the engine explains itself to your copilot

Every Threenity project ships a docs/threenity/ folder — a full capability inventory, scripting API reference, and per-pack authoring guides, kept in sync with the engine version. Claude Code, Cursor, Copilot, or any other assistant reads the real, current API from inside the repo — no hallucinated methods from a two-year-old blog post.

Files are the API

Agents author by editing files — components, scenes, assets, project config — and the engine is a robust reader that applies changes safely. There's no fragile UI-automation layer to script against, which is precisely how coding agents already work best.

The open 3D slot

2D web engines have started marketing themselves as AI-ready. For 3D on the web, the combination — a library models already know (Three.js) plus an engine architecture designed for machine authoring — is the bet Threenity is making. Judge it by the design decisions above, not the label.

Go deeper

• The architecture underneath: a real game engine for Three.js.
• What humans get: the visual editor.
• Back to the Threenity homepage.