CAD engineers can’t sleep after seeing this
I watched the demo of text-to-CAD, an open-source project that generates a complete 7-DOF mechanical arm from natural language, complete with kinematics, a custom GUI, and then exports to STEP, slices it, and sends it directly to a printer.
I used to think this was just something to mess around with. I didn’t expect it could actually output usable STEP files.
Let me share a real feeling: the most torturous part of building a hardware prototype isn’t drawing the diagrams itself—it’s the wall between CAD and code.
After finishing the model, you have to manually export URDF, write coordinate transforms, tweak a bunch of parameters, and a single person ends up juggling six or seven tools. One wrong bump in the pipeline and it crashes. The intermediate format “agent” produces can’t be read at all.
This harness turns every step—modeling, exporting, describing, slicing, and sending—into functions that an agent can call.
With Claude Code and Codex as the orchestrators, you only say one line: “I want a mechanical arm that can grab things,” and it handles the rest.
The generated geometry is probably rough for now, but don’t forget its iteration speed.
Previously, changing one joint position meant: open SolidWorks, edit the sketch, update the assembly, re-export the URDF, and re-compute kinematics. Now you change the prompt and, ten seconds later, you can see a new model. That broken pipeline gets killed by a file.
What makes it feel even more worth it is that it runs locally. It doesn’t depend on the cloud, and iterating through trial and error doesn’t come with API-cost anxiety. Developers should be willing to pay for this kind of freedom.
Unclear part: with complex assemblies that include gears, tolerances, and cables, can it really be done well with pure prompts?
But the starting point is already far lower than I imagined.
#AI硬件 #开源工具 #CAD