Claude can actually do CAD now in @Onshape
Here it worked for an hour and built a 4-part monitor arm, starting only from a sketch and description. The trick was to give it the tools to look at its own work.
Introducing: Jarvis Onshape MCP
Feature request: It would be trivial to make this popup list the tapping config and give forgetful people like me one last chance to review.
@jimbelosic
Claude can actually do CAD now in @Onshape
Here it worked for an hour and built a 4-part monitor arm, starting only from a sketch and description. The trick was to give it the tools to look at its own work.
Introducing: Jarvis Onshape MCP
Claude can actually do CAD now in @Onshape
Here it worked for an hour and built a 4-part monitor arm, starting only from a sketch and description. The trick was to give it the tools to look at its own work.
Introducing: Jarvis Onshape MCP
@MikushRab Awesome work, the NIST dataset is hard! If you haven't yet, try Model Mania, I found that it's a bit easier for the agent. What kind of info is in the rendered validation?
@Orwelian84@Onshape Conversation: 467K, but real usage was 85M Tok. Every turn re-reads the prefix cache, so 331 turns got VERY hungry. Even on Max 20x, it's a lot. Notice the input is only 688 tokens, that's what I actually gave the model, the rest is automatic
@Principal_Jon@Onshape Yes! Would recommend you try and run this, or something like ADAM CAD (same idea as this project). It makes 3D-print-ready files
@JGoodwillieV@Onshape Yeah it's pretty token-hungry. Lots of images being used for renders, and the work can take lots of turns which eat up cache context tokens
@coreyward I'm planning on it! Gemini's tool use is less advanced so trying to combine Claude's tools and Gemini's vision. I use Gemini for all my CAM advice already. More coming soon...
@CRudinschi@Onshape Yesterday we used it to make a two-part camera mount, which it got right in one prompt and we then 3D printed and used.
It's still highly dependent on vision model capability to succeed, but the CAD feature complexity is the "easy" part
Full writeup with benchmark data: I tested Claude on a dozen reference parts. The (un)surprising result is that the bottleneck isn't CAD execution, it's image understanding. Once you describe a part cleanly, the model builds it reliably.
https://t.co/Q7GJcjtwS2