67.js: 3D Engine Driven by AI Agents

What changed

67.js is a WebGL2 renderer that has been enhanced with a suite of addons to support various 3D elements, including terrain, vegetation, characters, liquid, and media export. The engine's core concept is to represent a scene as plain data, where each component has geometry, a transformation matrix, and material properties. A key feature is its design for AI agents to construct and modify scenes through an MCP (Model-Context Protocol) server. Agents can interact with the system by viewing PNG previews, rendering still images, or generating MP4 videos. The project explicitly states its goal to match the capabilities of Blender while maintaining the lightweight nature of three.js. The latest release, v0.18.0, introduces fresh updates and signals for AI and developers.

Installation is streamlined for AI agents, with prompts provided for popular coding assistants like Claude Code, Codex, Cursor, Gemini CLI, and OpenCode. These prompts guide the agents to install the package via npm (@codepawl/67js) and connect to the MCP server. For manual installation, the package is available on npm as 67js@npm:@codepawlhq/67js. The project requires Node.js 20.19 or later for its tools, and rendering necessitates a WebGL2-compatible browser.

The engine's architecture allows for dynamic scene updates through renderer.updateParts(parts), enabling modifications to geometry or materials without full scene reloads. It also supports responsive canvas resizing via renderer.resize() and proper cleanup with renderer.dispose(). The MCP server component requires a local Chrome installation and can be run with npx 67js-mcp, with options to specify the backend (gpu or cpu) and memory allocation. Agents interact with the MCP server by authoring JSON recipes, defining lights and fog, importing GLB assets, and specifying animations. The server is designed with security in mind, accepting no arbitrary code or remote asset URLs. Two specific skills, build-67js-scene and render-67js-media, are provided to teach agents how to manage scenes and export media.

67.js offers an extensive list of addons, categorized by functionality. These include advanced features for lighting (HDRI), terrain generation, vegetation scattering, character rigging and animation (glTF import, retargeting, IK), liquid simulation, vehicle dynamics, physics integration (Rapier), live game development (runtime, input, networking, spatial audio, animation graphs, player controllers, game rules), and modeling tools (mesh editing, sculpting). For rendering, the engine supports exporting images and videos using headless Chrome, with options for GPU or CPU backends and various quality presets.

Why it matters for builders

This project offers a powerful, agent-controllable 3D rendering environment, bridging the gap between AI capabilities and visual content creation. Builders can integrate 67.js into their AI agent workflows to automate the generation of 3D assets, scenes, and animations. The extensive addon system provides flexibility for creating diverse visual experiences, from realistic terrains to complex character animations, all driven by AI logic.

Practical impact

Developers can explore integrating 67.js with their preferred AI coding assistants by following the provided installation prompts. Experiment with the quick-start example to render basic scenes and then delve into the documentation for specific addons like terrain or character animation. For AI agent developers, the MCP server and associated skills offer a direct pathway to programmatically control 3D scene creation and media export, enabling automated content pipelines for games, simulations, or visual storytelling.

Caveats and source limits

The source mentions performance claims in verification.md but does not provide specific benchmark numbers or direct comparisons within the provided text. While the goal is to be as capable as Blender, independent verification of this claim is not detailed. The project is described as a "fresh release" with version v0.18.0, indicating it is actively developed but may still be subject to ongoing changes and potential instability. The MCP server requires a local Chrome installation, which might be a dependency constraint for some deployment environments. The source does not specify pricing or licensing beyond the MIT license for the core library.