AI and Onshape, how to get around the API quota
So I'm not sure how many folks have tried out using AI for CAD.
With the newer models Sol 6.1/Astra 6 or Opus 5.5 it is actually quite capable.
One of the best uses I have discovered is taking a step file (say gobilda's kitbot or the itkan bot) and having the AI organize all of the parts, make all of the mate connectors, etc. This woudl take hours to do by hand but is done easily by the AI using Onshape or Fusion. Here is an example of a cleaned up itkan robot step file with proper mate connectors added by AI: https://cad.onshape.com/documents/912406ea9713e41a95a75a19/w/9027348954e2d2ededee50f1/e/79ce681fbe5a7113ed9af07f?renderMode=0&uiState=6ac65fe6f580da0845bdb980
or how we used AI to take the gobilda starter bot and make the plastic bits 3d printable (we owned most of the parts, but didn't have any of their polycarbonate sheets: https://cad.onshape.com/documents/40e414fdf67a27907e5a9743/w/f84008ed7fd239c637ba2ae8/e/dcf5368543317b415ce57e25
It also isn't too shabby with making specific parts. Here is a turret design with a movable inner hood done entirely with prompting (no human has ever directly touched this file). It uses a servo mounted cam to move the inner movable hood to swap between pollen/nectar and the hood runs on some linear bearings/rods. The turret itself uses a very large bearing encased in 3d printed parts. We probably aren't going to use this design but we did it to see just how far AI can go (turns out pretty far):
However one of the problems with using AI assisted CAD with onshape is that they are very strict with API limits (with the official Onshape MCP). You get a YEARLY (not weekly) allotment and you can easily burn through the API quotas ia a week (you only get <10 API calls per day as I think they give you 2500 api calls per year). I burned through half of my yearly alottment in 2 days before I figured out that such just wasnt going to work.
The way around this is to make the AI agent use a signed in session (not the API key) using the API Explorer. This has rate limits but not yearly limits. This does mean you will have to sign in your AI's browser every session (as it can't use a stored API key) but you can see from the examples above it works pretty well.
To instruct an agent to do this right and to also give it access to local caching/be more efficient we came up with our own local onshape MCP server along with some agent rules (for example we do instruct it to use a local instance of build123 for rapid prototyping before creating native onshape files). It also is able to pull parts from the FRCDesign tool (though often the agent just goes and downloads step files directly from manufacturers pages).
https://github.com/ARES-23247/ares-onshape-local
Not sure this will be useful to others or not but we have been pretty impressed with the rapid improvements in the most recent models.
Of note Fusion has a built in MCP server with a lot better AI support and doesn't require this, but, as a team, we are still wanting to use onshape. Again you cannot use the Onshape MCP server as the yearly limit is very very low with the api key. You have to use the API Explorer work around. Our ares-local-onshape mcp server helps a bit by caching things and the skills a bit by encouraging agents to use build123 or other local lightweight CAD programs to so modeling (figuring out intake wheel spacing etc) before doing things in the cloud.
Probably going to get downvoted by the AI haters but the truth is nobody is going to be spending hours cleaning up mate connectors, etc. in the future and it is pretty clear (just like with programming) AI is going to be increasingly integrated into workflows.
