Download for Windows  ·  Pitch deck  ·  How it works  ·  90-second demo  ·  Architecture

---

What it does

Describe a part in plain English, or in terse typed commands, and get an exportable STEP / STL / engineering-drawing PDF in seconds. The geometry is real B-rep, so you can hand it to a CAM package or a machinist. Then mesh and solve it in the same tool.

No API key required: it runs in typed-command mode out of the box, or against a free local Ollama model with no network at all.

The loop

Most AI CAD demos stop at "looks like a part". The difference here is what happens after the geometry exists: five agents inspect it, and then the same geometry gets meshed and solved.

The visual critic actually looks at a render through a vision model. The DFM critic catches "wall too thin" before you ship, not after. The standards critic reads the notes you saved once and applies them to every part after that.

What it can build

Why bother

Real B-rep, not mesh triangles	Output goes to CAM or a machinist. Mesh-only tools cannot.
Critics run before export	A design loop that rejects its own bad output is worth more than a faster generator.
Bring your own LLM	Anthropic, Gemini, local Ollama, in-browser WebLLM, or none at all.
Simulation on the same geometry	Linear-elastic and steady-state thermal, on the B-rep you just made, not a re-import.
Drawings, not just models	4-view PDF with dimensions, title block and mass properties.
It remembers your rules	Saved notes are retrieved into every agent prompt with TF-IDF.

Quick start

Windows

1. Download ChatCAD_Setup.exe
2. Double-click. One-time Miniforge and CadQuery install, about 5 minutes and 1.5 GB.
3. Launch from the desktop shortcut. Your browser opens at http://127.0.0.1:5000/.

From source

conda create -n chatcad python=3.11 -y
conda activate chatcad
conda install -c conda-forge cadquery -y
pip install -r requirements.txt
python app.py

Pick an LLM (optional)

Backend	How	Cost
Anthropic Claude	paste an sk-ant-... key in settings	best quality, about half a cent per part
Google Gemini	paste an AIza... key from aistudio.google.com	free tier
Local Ollama	install Ollama, then ollama pull qwen2.5	free, fully offline
In-browser WebLLM	pick a browser: model from the dropdown	free, no key, 600 MB one-time
None	typed-command mode	works out of the box

Keys are sent with the request and are not stored server-side. They live in your browser's localStorage.

90-second demo

1. Launch                       clean viewport, feature tree, view cube
2. bolt_stack demo M8 15 60     plate + 2 washers + threaded M8 bolt + hex nut
3. right-click bolt             Mirror about XY, second bolt appears
4. Render tab                   Polished steel, apply to all
5. File tab                     Drawing PDF, 4-view drawing downloads
6. Simulate tab                 click the plate, Run FEA, real stress in 8 s
7. Design Agent mode            "turbojet engine, 200 mm fan, 600 mm length"
                                19 sub-parts build themselves, critic checks
                                each milestone
8. Render tab                   Brushed aluminium + outdoor sky, catalog shot

Architecture

flowchart TD
    U["Chat or typed command"] --> A["app.py<br/>Flask: chat, scene, drawing, FEA"]
    A --> AG["agents.py<br/>planner, modeler, 3 critics"]
    AG --> L{"LLM backend"}
    L --> L1["llm.py<br/>Claude tool-use"]
    L --> L2["llm_gemini.py"]
    L --> L3["llm_ollama.py<br/>offline"]
    AG --> CE["cad_engine.py<br/>CadQuery / OpenCascade"]

    CE --> S["sketch_engine.py<br/>2D sketcher, scipy constraints"]
    CE --> AS["assembly_engine.py<br/>mate solver"]
    CE --> LB["library.py<br/>fasteners, gears, bearings"]
    CE --> PR["profiles.py<br/>T-slot, I-beam, channel"]
    CE --> SM["sheet_metal.py<br/>bends, flanges, boxes"]
    CE --> MA["materials.py<br/>density, mass properties"]
    CE --> RC["assemblies_recipes.py<br/>turbojet, gear_train"]

    AG --> KN["knowledge.py<br/>TF-IDF RAG over your notes"]

    CE --> V["templates/index.html<br/>Three.js viewport, PBR + IBL"]
    CE --> D["drawings.py<br/>4-view PDF"]
    CE --> IO["step_io.py<br/>STEP in and out"]
    CE --> F["fea.py + fea_worker.py<br/>gmsh + scikit-fem"]

    V --> OUT["STEP · STL · PDF · stress plot"]
    D --> OUT
    IO --> OUT
    F --> OUT

Loading

Pricing (intent)

Tier	Price	What you get
Open source	$0	Full feature set, runs locally, MIT
Pro (planned)	$49/mo	Hosted instance, priority support, custom knowledge base
Team (planned)	$499/mo	Shared knowledge base, multi-user sessions, audit logs
Enterprise (planned)	custom	On-premise, SAML/SSO, domain fine-tune

What this is, and what it is not

It is a chat-driven CAD tool with real B-rep output, real FEA, and real engineering drawings. Comparable in spirit to a text-to-CAD demo, with broader operation coverage and a multi-agent design loop on top.

It is not a SolidWorks, Onshape or Fusion 360 replacement. Those kernels carry 25 to 35 years of customer-validated edge cases. This does not have that history. The wedge here is speed from a chat prompt to a serviceable part, not authoring 50,000-part assemblies with full tolerance stack-ups.

It will be whatever the first ten paying customers say it needs to be.

License

MIT. See LICENSE.

---

Built by Samarjith Biswas, Ph.D.

Mechanical engineering  ·  acoustic metamaterials  ·  agentic AI for design