Your AI subscription, now a CAD intern for 3D printing
nurb lets you describe a part in plain English, then has your existing AI model it, check it against print physics, and hand you a slicer-ready file.

What it does
nurb is an agentic CAD layer that sits between the AI you already pay for and your 3D printer. You describe a part in conversation; the AI writes a Python function using build123d on the OCCT kernel, and nurb compiles it into a true B-rep solid. It then runs thirteen printability checks against the exact geometry—overhangs, thin walls, floating islands, warp risk, stability—and feeds the result to a live viewer with sliders for every parameter. When you are satisfied, it exports a print-ready 3mf tuned to your printer.toml profile, or hands off to OrcaSlicer or Bambu Studio for time and cost estimates.
The interesting bit
The tool treats the AI as a fallible junior designer rather than an oracle. Because the AI cannot see, nurb gives it eyes: printability findings pin themselves to exact faces on the model, and real-world dimensions from phone scans or downloaded meshes are enforced in measurements.toml so the AI cannot guess them. The stress simulation even quotes failure margins against layer adhesion, which is where FDM prints actually fail, not some abstract isotropic material science.
Key highlights
- True B-rep solids on the OCCT kernel, not meshes, so chamfers and fillets are real operations and STEP export opens cleanly in Fusion or FreeCAD.
- Thirteen physics-aware checks run against the exact solid, including overhang angles, bed adhesion warp risk, and center-of-mass stability, with findings mapped to specific faces.
- Live viewer updates without moving the camera, exposes a slider for every parameter, and estimates print time and filament cost by driving your existing slicer.
- Works fully offline: modeling, checking, viewer, and export all run locally with no account or cloud dependency.
- Stress analysis uses voxel simulation to show load concentration and sag, quoting break margins against layer adhesion rather than idealized bulk strength.
Caveats
- The first build takes about 45 seconds while the CAD kernel loads cold; subsequent edits land in under a second.
- The
min_wallcheck samples faces, so it can miss a thin pinch that no ray happens to hit—a clean report means “none found,” not “none exist.” - A hosted configurator and in-viewer measurement tools are listed as not yet built.
Verdict
Grab this if you want to design functional 3D-printed parts by conversation and need physics-aware validation without touching cloud CAD. Skip it if you need a hosted configurator, in-viewer measuring tools, or a geometry kernel that loads instantly.
Frequently asked
- What is Shpigford/nurb?
- nurb lets you describe a part in plain English, then has your existing AI model it, check it against print physics, and hand you a slicer-ready file.
- Is nurb open source?
- Yes — Shpigford/nurb is an open-source project tracked on heatdrop.
- What language is nurb written in?
- Shpigford/nurb is primarily written in Python.
- How popular is nurb?
- Shpigford/nurb has 516 stars on GitHub.
- Where can I find nurb?
- Shpigford/nurb is on GitHub at https://github.com/Shpigford/nurb.