I’ve been using McNeel’s Rhino MCP server and Grasshopper 2 (Rhino 9 beta) with Claude, and it’s been going surprisingly well.
My work is mostly product design (extruded or 3-axis machined parts) plus product rendering. I know an experienced user can model this faster by hand. This post is about what the AI workflow is like and what makes it work better.
Task 1: Rendering, building and tuning PBR materials and lighting
I gave it a photo of a finished product and had it match the camera and lens to the shot. Then it iterated on lighting and materials until the render was close to the photo.
- Ask for a spread of options (A/B/C/D) and choose yourself. It isn’t good at judging what looks “real” in fine detail.
- Give it reference photos of the real material if you have them.
Task 2: Rebuilding a STEP part as a parametric GH2 definition
I set this up as building a reusable skill that:
- Asks how the part will be manufactured (in this case an extruded profile, machined on four sides). That decides how it gets sketched.
- Rebuilds the part in GH2 with named sliders. A GH2 graph is much easier for an AI to work with than raw geometry, because it can change one value and re-solve.
That worked well. The pain point was making changes. To adjust a single fillet on a baked part, I had to trace it back through the GH definition myself and then tell the AI where to look.
So, I (As in asked Claude to) build a small tool called FeatureContext:
- Bake with provenance: Parts are baked with GH2’s own bake, which tags each object with the component it came from, plus some extra metadata.
- Pick and comment: You pick a face or edge in the viewport and type a note, e.g. “make this a hard edge”. The script walks back through the GH2 graph to the component that created that face and lists the sliders that drive it. It copies a summary to paste into the AI chat and saves a screenshot with the face highlighted.
With that, the AI knows exactly which component and slider to change, rather than making a long chain of tool calls to find it.
Questions
- Does something like this already exist? As far as I can tell, the context you can currently give the AI is the viewport, layers, or whole objects.
- McNeel folks: is feature-level context like this on the roadmap for Rhino MCP / the AI tools?
