Amagine3D turns a hardware brief into an editable BRep enclosure, and the repo has 4,578 stars in 26 days

The organisation behind amagine.ai has published Amagine3D, an open-source tool that takes a hardware description, reference images, and key dimensions, then designs an enclosure and the assembly structures around it. The repository was opened on 19 August 2026 and has picked up 4,578 stars and 205 forks in 26 days. It is Apache-2.0 licensed.

Editable BRep solids and a real STEP file

Amagine3D produces editable build123d Python source, a genuine STEP file, an STL, and a 3MF package for multi-colour printing. The project says every manufactured part stays a valid BRep solid, so the geometry is the kind an engineer can carry into Fusion, SolidWorks, or a slicer without a repair pass first. Parameter changes are written back to the source, and the workbench rebuilds the geometry from the new numbers.

The README ships one worked example: a desktop enclosure for the BUSY Bar productivity device, generated from public information and reference images. The tool placed the display window and controls, split the shell into printable parts, and left the appearance and assembly dimensions as editable parameters.

The design state is what decides the next action

Amagine3D runs a two-stage loop. An inner autonomous loop reads the current design, plans a change, runs the change in a real geometry kernel, and reads back measurements. Wall thickness on a cavity is checked against the actual solid, so the number a section inset would give is ignored. Assembly interference, motion paths on hinges and sliding covers, and exported-file readback are all checked before a candidate becomes the new accepted version. The project calls this a "3D-native Agent", and the current design state is what decides the next action.

The stack underneath is public open source. Geometry runs on Open CASCADE through CadQuery OCP and build123d. Manifold and trimesh do mesh boolean checks. lib3mf writes the 3MF packages. Three.js drives the browser viewer. OpenAI Codex is the agent runtime, and the local Express server streams SDK events into a stable contract so the app does not import Codex types directly. Each session gets its own workspace and Codex state directory.

Runs on your laptop with a model gateway

Amagine3D is a local Node and Python install. Node 22.19.0 or newer, Python 3.10 through 3.13, and an OpenAI Responses-compatible API key or model gateway. No desktop CAD application is required. The workbench runs at 127.0.0.1:6160 and the local API at 127.0.0.1:6161. Codex sessions run with workspace-write and no UI approval inside the session directory, and writes outside it stay sandboxed.

The project describes itself as under active development, focused on single- and multi-colour printable BRep geometry, and tested on desktop Chrome and Edge. Surface continuity can be coarse while the design keeps valid solids and manufacturing checks.

Open-source AI-native CAD has been a thin category until now. A tool that returns editable source and a genuine STEP file is worth trying against real enclosure work: for a small hardware team, that is the difference between an AI output that starts a project and one that ends at a screenshot.

Source

Amagine3D on GitHub: github.com/amagine-ai/Amagine3D.