Rhino MCP — AI-Assisted Architectural Modelling in Rhino 3D
A Model Context Protocol (MCP) server that gives Claude, ChatGPT, Codex, Gemini or a local Ollama model full control of Rhino 8 / Rhinoceros 3D. Describe a building in plain language and watch it get modelled — massing, floor plates, cores, facades, vaults, plans, sections and area schedules — with the AI checking its own work as it goes.
Rhino MCP server: AI-assisted architectural modelling in Rhino 8 with Claude, ChatGPT, Codex, Gemini or local Ollama. 128 tools: massing, vaults, arches, drawings, materials, PDF tracing, plus intent validation that catches errors valid geometry hides. Protocol 5.1, no .NET SDK required.
ดาว
32
เครื่องมือ
—
อัปไทม์
—
Built by an architect, for architects. No coding required to use it.
Keywords: rhino mcp · rhino 3d mcp server · rhinoceros mcp · model context protocol rhino · rhino 8 ai plugin · claude rhino integration · ai architectural modelling · ai 3d modelling · generative architecture · computational design · rhinocommon automation · architectural bim ai
Why this Rhino MCP?
Several Rhino MCP servers exist. This one is built for long, real modelling sessions — one user drove it to ~8,000 objects reconstructing a Gothic cathedral at 1:1, with zero invalid breps.
This project
Tool surface
128 tools, curated into lean / standard / full profiles so small models aren't drowned
Transport
Protocol 5.1 — multiplexed, so reads, ping and cancel answer instantly while a 3-minute script runs, and long batches report live progress
Reliability
Idempotent retries (no duplicate geometry on reconnect), atomic batches with rollback, write-ahead-log crash recovery
Correctness
Intent validation — asserts what the geometry means, not just that it parses
Geometry stdlib
rab helpers auto-imported into every script: walls, slabs, grids, pointed arches, rib vaults, rose tracery, mouldings
Vision loop
Auto-thumbnails after every edit, multi-angle review sets, before/after pixel diffs, one-call section previews
Install
Pre-built plugin — no .NET SDK needed, one double-click, and it verifies itself
If Rhino MCP saves you modelling time, please star the repo — stars are how other Rhino users find it.
What makes it different: the AI checks its own work
Valid geometry is not correct geometry. An AI generating thousands of parametric objects makes arithmetic and wiring errors — a doubled base height, swapped arguments, a boolean cutter added instead of subtracted — and every one of those produces a perfectly closed, valid solid that validate_objects would happily pass.
section_preview — cut the model at any station and look inside, in one call, leaving no geometry behind
validate_objects — separates real corruption from intentionally open shells, with naked-edge lengths so a hairline gap stands out
For architects — no coding needed
If you can describe a building, you can model it:
"Create a 4-storey office massing on a 30×18 m footprint, taller ground floor, derive the floor slabs, add a core with two lifts and a stair, then put windows on the north facade every 3 metres."
You can ask for:
Massing studies — footprints, levels, setbacks, options side by side
Floor plans and sections — cut, captured and restored without touching your viewport
Area schedules — GFA/NFA by layer, level or name, ready to paste into a report
Historical work — pointed and equilateral arches, rib vaults, rose windows, mouldings
Materials and views — "make the facade glass, the core concrete, show me a hero shot"
Drawing imports — trace a PDF site plan or import a DWG and build on top of it
Quick start
Download this repository (green Code button → Download ZIP, then extract — or git clone).
Close Rhino and your AI client, then double-click INSTALL.bat.
It installs the plugin, the Python server and configures every AI client it finds, then prints a PASS/FAIL verification report and writes install-log.txt. It will not close without telling you what happened.
Open Rhino 8. First time only: run PlugInManager, click Install, and choose
%APPDATA%\McNeel\Rhinoceros\8.0\Plug-ins\RhinoAIBridge\RhinoAIBridge.rhp.
After that the bridge starts automatically with Rhino — no command to type.
Restart your AI client and ask it: "ping Rhino".
Requirements: Windows, Rhino 8, and one MCP-capable AI client. No .NET SDK required.
Access modes are set with the RHINO_AIBRIDGE_MODE environment variable — safe (blocks code and destructive edits), standard (allows deletes/booleans), or developer (everything, required for execute_python3).
The rab geometry library
Every execute_script call gets rab auto-imported, so the model writes intent instead of boilerplate:
The two-centred arch solver uses c = (h²−s²)/2s, R = c+s, and the extrados is concentric (same centres) — which is what makes archivolt offsets exact. Vault webs blend to a flat fold at the boss so they land precisely on the crown, where a naive Coons patch overshoots.
Reusable code:write_module(name, source) saves a library that later scripts import with rab.use('name'). The server writes the file, so nothing inside Rhino ever holds a file handle open.
Tool surface: 128 tools, three profiles
Set RHINO_TOOLS=lean|standard|full in the server environment:
Profile
Tools
For
lean
21
small/local models (Ollama), minimal context
standard(default)
76
Claude / GPT-class daily driving
full
128
everything, including JSON twins and compatibility aliases
Anything not exposed in the active profile is still callable as a batch sub-command, and the live command list is always available from list_commands and the rhino://capabilities resource.
Scene & context
Tool
What it does
ping
Health check — Rhino version, units, scene_version etag, and which script engines are actually available
query_scene
Universal query: filter by type, layer, name, bbox; columnar format saves 40–60% tokens
list_commands
Live plugin dispatch table (never drifts from the code)
skills/rhino-architect/ is an Agent Skill that teaches the model how architects work: phase ordering (brief → layers → massing → verify → structure → envelope → drawings), millimetre defaults, when to batch vs. step, verification cadence, anti-patterns, and debugged parametric generators for stairs and curtain walls. Point your client at that folder, or install the packaged .skill.
Connecting AI providers
INSTALL.bat configures whatever it detects. Manual configuration:
Protocol 5 multiplexes by request_id, so reads and ping answer from the plugin's socket thread while the UI thread runs a long script. Mutating requests are registered for idempotent replay: if the connection drops after delivery, re-sending the same request_id replays the cached result instead of duplicating geometry. Viewport captures travel as raw binary frames rather than inflated base64.
Atomic batches run inside a single Rhino undo record; any failure triggers one Doc.Undo().
Checkpoint economics: snapshots are skipped when the scene hasn't changed, throttled on large documents, and controllable per call with checkpoint="off" | "auto" | "force".
Building from source
Only needed to modify the C# plugin. Target machines never need the .NET SDK.
cd plugin && dotnet build --configuration Release
cd ../server && uv sync --group dev && uv run pytest -q && uv run ruff check src tests
CI builds the plugin, lints, and runs the test suite on every push.
Troubleshooting
"Cannot connect to Rhino" — open Rhino. The bridge auto-starts; if it doesn't, run AIBridge once and re-check PlugInManager shows the plugin as loaded.
The installer closed instantly / I don't know if it worked — fixed in v4.11.0. The installer now prints a verification report and always waits for ENTER. Every run writes install-log.txt; attach that to an issue.
"Rhino is running" / "MCP server is running" during install — both hold their files open. Quit Rhino and your AI client, then re-run.
Tool not found in my client — check your profile. RHINO_TOOLS=standard hides ~50 long-tail tools; they remain callable via batch, or set RHINO_TOOLS=full.
execute_python3 unavailable — call ping and read script_engines.python3.reason. It states exactly what's missing (Rhino 8.11+, the rhinocode CLI, or Developer mode).
Health check — powershell install/rhino-mcp-healthcheck.ps1, or cd server && uv run python ../scripts/doctor.py.
Changelog
v4.16.0 (current)
Live progress for long operations (protocol 5.1) — batches report op 7/40 as they run, and long scripts, area schedules and clash checks send a heartbeat (running 42s of 180s budget), all surfaced as MCP progress notifications, so a long build no longer looks like a hang. Negotiated: older plugins and clients keep working unchanged
execute_python3 fixed — it now returns the script's real stdout/stderr, reports a script that raises as an error with a traceback (it previously came back ok), and actually honours timeout_seconds
Session distiller (evals/distill_session.py) — turns the write-ahead log every session already keeps into a report (command mix, failures, retries, slow calls) and a draft eval task measured from the live scene
The dist/ staleness check now also verifies the protocol version and advertised features — it previously passed while dist/ lacked protocol 5.1 entirely
Confirmation before big deletes and checkpoint restores (MCP elicitation) — the server knows what the model doesn't: that delete_objects(["all"]) means 842 objects including hidden ones, or that a restore throws away everything since the save. Clients that support elicitation ask you first, quoting the real count; saying no means nothing is touched. Clients without it behave as before
Distilled eval tasks check elevation, not just size — a roof dropped from 8 m to the ground used to pass every drafted assertion
build.bat fixed: it no longer reports BUILD SUCCESSFUL when Rhino blocked the install (it also used to delete the plugin's DLLs before failing), and it now works from any folder. It also reads its version from VERSION; release notes for v4.10–v4.15 added in docs/
checkpoint="off" now works on execute_script — it was silently ignored there (as was "force"), so read-only scripts still wrote a full .3dm snapshot, the cost the setting exists to avoid on big models
ping reports what your MCP client supports — whether it can show confirmations (elicitation) and whether it asks for live progress, so you can see why a feature is or isn't visible instead of guessing
v4.12.0 – v4.15.0
Self-describing tool schemas, capture(fit=...), layer/selector fixes, assert_dimensions, capture_elevations, run_selftest and eval task 10 — see docs/RELEASE_NOTES_v4.10-v4.15.md
v4.11.0
Installer rewritten — verification report, install-log.txt, detects locked files from a running Rhino or AI client, and never closes without explaining itself
.gitattributes pins Windows scripts to CRLF — LF endings made cmd.exe mis-seek on goto/call, which is why the installer appeared to close instantly for ZIP downloads
Shipped plugin refreshed — dist/plugin/ had drifted a month and a half behind the source
CI now fails if dist/ is out of sync with a fresh build
Issues and pull requests welcome — especially field reports. The most valuable contribution this project has received was a detailed write-up of an 8,000-object modelling session, which produced seven confirmed bug fixes and an entire new category of tooling. If you push this thing hard and it bends, please tell us how.
License
MIT — see LICENSE. Free for personal and commercial use.
Rhino MCP server: AI-assisted architectural modelling in Rhino 8 with Claude, ChatGPT, Codex, Gemini or local Ollama. 128 tools: massing, vaults, arches, drawings, materials, PDF tracing, plus intent