What it does

Articraft generates articulated 3D objects—think desk lamps with hinged arms, fans with tilting heads—by having LLMs write Python code instead of manipulating polygons by hand. You prompt it, the system produces a model.py file, and you get geometry with semantic parts and physical joints. A local React viewer lets you browse, search, and inspect results.

The interesting bit

The project treats 3D generation as a software engineering problem: versioned records, forking existing assets with natural language edits ("make the handle longer"), and Git LFS for lazy-loading heavy payloads. It also explicitly supports external agents—Claude Code, Codex, Cursor—as first-class citizens, not just API fallbacks.

Key highlights

• CLI-driven workflow: articraft generate, articraft fork, articraft data hydrate
• Multiple LLM backends: OpenAI, Gemini, Anthropic, DashScope/Qwen
• Image-conditioned generation from reference photos
• Dataset contributions released under CC-BY 4.0 for ML training
• Apache-2.0 licensed; Python 3.11/3.12, uses uv and just

Caveats

• Generated model.py files are executed as Python code; the README warns to only run trusted sources
• Python 3.13+ not supported
• Requires uv, just, Git LFS, and optionally npm for the viewer—no lightweight pip install

Verdict

Worth a look if you’re building 3D datasets for robotics, simulation, or generative ML and want programmatic, versioned assets. Skip it if you need polished, production-ready models out of the box; this is a research pipeline with rough edges.