An open-source project turns Claude Code and OpenAI Codex into a CAD engineer, generating editable 3D models from plain English. The tool exports STEP, STL, DXF, GLB, and 3MF files, plus G-code for CNC and 3D printing.
Key facts
- Exports STEP, STL, DXF, GLB, 3MF formats
- Generates Robot URDF and G-code
- Direct Bambu Lab integration
- Runs 100% offline, fully open source
- Demo builds 7-DOF robotic arm from prompts
The era of "I can't design CAD" may indeed be over. A new open-source project, announced on X by @_vmlops, taps into the code-generation capabilities of Anthropic's Claude Code and OpenAI's Codex to produce real, editable 3D models from natural-language descriptions According to @_vmlops.
The tool supports a wide range of export formats: STEP, STL, DXF, GLB, and 3MF. It also generates Robot URDF files for robotics simulation, G-code for CNC machining and 3D printing, and includes direct integration with Bambu Lab printers. All processing runs 100% offline, and the codebase is fully open source.
One demo showcases the system building an entire 7-DOF robotic arm — kinematics, CAD model, and GUI — generated mostly from prompts. This is a significant step beyond prior text-to-CAD efforts, which often produced non-editable meshes or required cloud APIs.
The unique angle here is not just the capability but the architecture: by using Claude Code and Codex as the reasoning engine, the project offloads the heavy lifting of geometry generation to frontier LLMs, while keeping the output pipeline local and open. This contrasts with proprietary offerings like Autodesk's Forma or Onshape's FeatureScript, which lock users into ecosystems. The offline, open-source nature also makes it viable for defense, manufacturing, and other sensitive environments where cloud data exfiltration is a concern.
However, the project is in early stages. The quality of generated models for complex assemblies or precision-tolerance parts remains unverified. The source tweet does not disclose benchmark results or comparison against human-designed CAD files — a gap that will matter for production use.
What this means for AI-native engineering
This project exemplifies a broader shift: AI models are no longer just generating text or code but are becoming the control plane for physical design tools. The combination of LLMs with open-source CAD kernels (like OpenCASCADE) could democratize mechanical design, much as GitHub Copilot did for software development.
For now, the most immediate impact is on prototyping and hobbyist workflows. Engineers can iterate on designs conversationally — tweaking dimensions, adding features, or exporting for fabrication — without touching a traditional CAD interface. The Bambu Lab integration suggests the tool is already targeting the desktop manufacturing market.
Limitations and open questions
The source provides no information on training data, model fine-tuning, or inference costs. It is unclear whether the system uses a single prompt or a multi-step agent loop. The 7-DOF arm demo is impressive but may not generalize to arbitrary mechanical systems. The project's GitHub repository was not linked in the tweet, so code inspection is not yet possible.
Key facts:
- Exports STEP, STL, DXF, GLB, 3MF formats
- Generates Robot URDF and G-code
- Direct Bambu Lab integration
- Runs 100% offline, fully open source
- Demo builds 7-DOF robotic arm from prompts
What to watch
Watch for the project's GitHub repository release and community benchmarks comparing generated models against human-designed CAD files. If the quality holds for precision parts, expect rapid adoption in prototyping workflows and potential acquisition interest from Autodesk or Onshape.
