Three.js, but better for an agent

ThreeNative keeps the Three.js programming model an agent already knows, then adds the engine, world-building and MCP toolchain plain Three.js leaves to you: 6 wired MCP servers, a live collaborative terrain editor, 200+ procedural assets, Unreal and Unity pack importers, deterministic playtests and native desktop/Android targets.

The floor is vanilla Three.js — zero hostage projects

ctx.scene is a real THREE.Scene and ctx.renderer is a real WebGPURenderer. There is no serialized scene format, no cloud-only runtime and no editor-owned state. Open your repo in any coding agent, ship MIT packages end to end, and drop down to raw Three.js anytime.

Less plumbing to reinvent

Plain Three.js makes an agent rebuild renderer setup, resize handling, the loop, fixed timesteps, input, physics, HUD sync, touch, gamepad, pause, test hooks and packaging. In the charter's audit of a vanilla game, about 42% of its roughly 400 lines was identical plumbing.

Read the guide

Six wired MCP servers, zero editor lock-in

Every scaffolded project wires 6 MCP servers (115 tools): offline engine discovery across 622 APIs, unified asset search with Unreal/Unity GLB conversion, collaborative Strata terrain editing, build-time GLB animation repair, 8-pass visual sculpting gates and live WebGPU DevTools.

Read the guide

The agent can play what it built

Playtest scenarios press keys, wait ticks and assert on game state, transforms, cameras, HUD, console errors, screenshots, triangles and draw calls. They fail closed, so a feature is checked by running the game, not by reading the diff.

Read the guide

Measured

Generic plumbing in a vanilla game
42%About 170 of roughly 400 lines in the abyss game were identical in every game.Source
4,096 cubes, native desktop
2.28 ms vs 4.17 msThreeNative shipped defaults against Godot, 1.8× faster. RTX 2080, 1280×720, median of 3 runs.Source
Native C++ engine, desktop
2.35× to 3.2× fasterFrame p50 against the previous host at three scene sizes. RTX 2080, same size and triangles. Web main thread: 1.41× faster.Source
512 skinned rigs
1026 → 4 drawsFrame time p50 47.05 → 24.45 ms against stock Three.js, faster in 6 of 6 paired runs.Source

Developer experience with an agent

These ratings are an assessment of how it feels to build with Claude or Codex, not a benchmark. The measured numbers are below, each with its source.

CapabilityThree.jsThreeNativeUnreal + MCP
Authoring loop
Ordinary TS/JS editing★★★★★★★★★★★★★★★
Claude already knows the API★★★★★★★★★★★★★★★
Direct Three.js access★★★★★★★★★★N/A
Changes are plain text and git diffs★★★★★★★★★★★★★★★
Vite dev loop★★★★★★★★★★N/A
Hot browser iteration★★★★★★★★★★★★★★★
Agent can grep the whole game★★★★★★★★★★★★★★★
No editor state to manipulate★★★★★★★★★★★★★★★
Verification
Deterministic automated gameplay tests★★★★★★★★★★★★★★★
Semantic scene inspection★★★★★★★★★★★★★★★
Performance assertions★★★★★★★★★★★★★★★
Visual verification★★★★★★★★★★★★★★★
Engine systems
Input abstraction★★★★★★★★★★★★★★★
Physics integration★★★★★★★★★★★★★★★
Asset pipeline + Unreal/Unity GLB import★★★★★★★★★★★★★★★
Procedural asset library & live terrain editor★★★★★★★★★★★★★★★
Build-time animation & sculpt MCP gates★★★★★★★★★★★★★★★
AAA engine tooling depth★★★★★★★★½★★★★★★
Targets
Browser★★★★★★★★★★★★★★★
Native desktop★★★★★★★★★★★★★★★
Android★★★★★★★★★★★★★★★
One source across targetsBrowser onlyYesMostly native Unreal
Working with an agent
Capability discovery★★★★★★★★★★★★★★★
AI-specific project instructionsManualBuilt inUsually custom
Token efficiency★★★★★★★★★★★★★★★
Autonomous feature implementation★★★★★★★★★★★★★★★
Recovery when the agent gets it wrong★★★★★★★★★★★★★★★
Parallel agents★★★★★★★★★★★★★★★
Agent-friendly architecture overall★★★★★★★★★★★★★½★

Browser-only libraries, codecs and DOM access limit what runs natively, and native packaging and device validation take longer than refreshing Vite. Unreal still leads on the depth of its mature AAA subsystems. Read the current challenges.