C33D

AI that designs editable 3D models — parametric CAD you can talk to.

Research Preview — works today, evolving fast. Try the live demo or run it locally in 4 commands.

Try it · How it works · Research · Contributing

---

What it does

C33D is a browser-based, AI-native parametric CAD system. You describe an object in chat, and an LLM agent architects it as a node graph. Because the output is a graph and not a static mesh, the model remains fully parametric — you can tweak dimensions, drag sliders, and remix features without breaking the design.

The underlying engine runs OpenCascade (via replicad) in WebAssembly — producing robust B-Rep solids exportable directly to STEP for manufacturing or 3D printing.

C33D uses a Bring Your Own Key (BYOK) architecture. API keys are entered once in the UI, stored exclusively in your browser's local storage, and never sent anywhere except the AI provider you chose. Nothing phones home.

Quickstart

git clone https://github.com/3esign/c33d.git
cd c33d
npm install
npm run dev

Then open http://localhost:5173 and enter your Gemini or OpenAI API key. Or use the hosted live demo.

How it works

1. Chat — You describe the object you want to build.
2. LLM Plans — The agent reasons about structural proportions and deconstructs the request into parametric operations.
3. Graph Generation — The agent emits a typed node graph (Primitives → Modifiers → Booleans → Compounds).
4. Validation Gate — The graph is schema-validated before execution; invalid connections are rejected outright.
5. Execution — A web worker runs the OpenCascade kernel in WASM to build the B-Rep geometry.
6. Geometry Percepts — The worker returns geometric reports (volume, bounding boxes, face counts) back to the LLM.
7. Auto-Repair Loop — If the geometry fails (e.g., self-intersecting fillet), the LLM uses the percepts to autonomously fix the graph.

What makes it different?

Feature	C33D	Grasshopper / Dynamo	CadQuery / build123d	Zoo (KCL) / text-to-CAD	Blender + LLM plugins
Paradigm	AI-authored node graph, human-editable	human-authored node graph	human-written code	AI-generated code/geometry	AI-driven destructive mesh edits
Output stays parametric	✅ sliders + formulas + relations	✅	✅	⚠️ code regenerable, no live UI	❌ mesh only
Real CAD kernel (B-Rep, STEP)	✅ OpenCascade	✅ Rhino	✅	✅	❌
Runs in browser, no install	✅	❌	❌	✅ (SaaS)	❌
Model-agnostic (BYOK, Ollama)	✅	—	—	❌ vendor-locked	varies
AI feedback loop (percepts + auto-repair)	✅	—	—	internal / closed	❌
Self-improving knowledge base	✅ human-gated	—	—	closed	❌
Evaluation of parametric integrity	✅ (unique metric)	—	—	❌	❌

The research angle

C33D is also an open research platform for parametric design intelligence:

1. The Knowledge Loop — The agent saves verified successful designs as few-shot examples, organically extending its own vocabulary based on human validation.
2. Evaluation Methodology — Unlike static text-to-CAD benchmarks (which measure if a shape compiles), our eval suite scores models on proportional integrity under parameter perturbation: does the design logic survive when slider values change? Does a car's wheelbase stretch coherently when you double the chassis length?

See the Research Directory for the foundational design documents (proportional coherence, sub-shape editing, vector/curve toolkit, node expansion, and eval upgrade plans).

Status & roadmap

C33D is a Research Preview — a working proof-of-concept for AI-native parametric modeling. The node library grows continuously. See docs/ROADMAP.md for what's next.

Contributing & citation

Contributions are welcome at every level — new nodes, eval runs, verified examples, or research RFCs. See CONTRIBUTING.md for the three contribution paths.

If you use C33D in your research:

@software{poturak_2026_c33d,
  author  = {Poturak, Semir},
  title   = {C33D: AI-Native Parametric CAD},
  year    = 2026,
  url     = {https://github.com/3esign/c33d}
}

---

C33D — Copyright 2026 Semir Poturak, PhD — Apache-2.0