I wanted to test how far the new official RhinoMCP could go if I put an AI agent in the driver’s seat for a whole design, not just one prompt. This is the result — a pavilion called Kanjun no Niwa (“the circling garden”): three nested rings of vertical louvers, entrance offsets, and a wave-form height variation.
The workflow was a relay:
- Concept (upstream): ChatGPT, image dialogue, to lock the form and a one-paragraph design intent.
- Implementation (downstream): Claude + the official RhinoMCP, writing the Grasshopper definition — built up in PoC stages (single ring → 3 rings via Data Tree → entrance offset + wave → Ladybug radiation extraction → Wallacei).
- Optimization: Wallacei, Pop 50 × Gen 30 = 1,500 individuals, each running a Ladybug annual-radiation analysis. Final D-balance case → 655 louvers baked.
A few honest notes for anyone trying RhinoMCP for real work:
- It handled Data Trees properly (graft/flatten for the 3 rings) — I expected it to stumble here; it didn’t.
- It drove Ladybug components (EPW → sun path → radiation mesh → threshold) end-to-end.
- The clean hard-stop: Wallacei’s run button. The agent wires Genes/Fitness correctly, but launching the solver UI and pressing run is still a human step — RhinoMCP can’t press in-plugin UI. Curious if others have found a way around interactive-solve UIs.
Full process write-up (free, with the failure modes): https://medium.com/@allelishi/i-let-ai-design-a-pavilion-from-a-3-line-prompt-to-655-louvers-via-a-chatgpt-claude-relay-8c0724b7435f
Built on Rhino 8 + official RhinoMCP + Claude Code. Happy to answer setup/wiring questions in the thread.