v2Open source · Rhino 8 · Windows & macOS

AI inside
Grasshopper —
you stay in control.

SmartHopper embeds AI in your definitions with native parametric components, an agentic canvas assistant, and a built-in MCP server — staged, undo-safe edits across 80+ AI tools and 8 providers. Free and open source: bring your own key, or run local models.

Start here →↗documentation↗github repo↗food4rhino
★ 30+GitHub stars↓ 1k+Installs8AI providers90+GH components80+AI toolsLGPLOpen source
Three ways to work
01

Native parametric components

Composable *2AI inputs and AI2* outputs embed AI inside your data trees — merge prompts, images, files and canvas state into a single call and get typed Grasshopper data back, no string parsing.

Text2AI input component feeding an AI to Text output component on the Grasshopper canvasscroll zoom · drag pan
docs ↗
02

Agentic canvas assistant

Click the AI button and chat with an assistant that sees your components, wires and errors. Every canvas edit is staged for your review — and undo-safe.

Talk to SmartHopper, your canvas assistant that can see what you are working onscroll zoom · drag pan
docs ↗
03

Built-in MCP server

Expose your canvas to Claude, Codex, Cursor or VS Code through a local, token-protected MCP endpoint — the same reviewed tool calls the assistant uses.

MCP Server component exposing a local streamable HTTP endpointscroll zoom · drag pan
docs ↗
Key features
/01

Native parametric components

Composable *2AI inputs and AI2* outputs embed AI inside your data trees — merge prompts, images, files and canvas state into a single call, get typed Grasshopper data back.

docs ↗
parametric_ai.gh*2ai → ai2*
AIPrompt, Img2AI and Text2AI components merging into AI to Textscroll zoom · drag pan
screenshot slot
/02

AI Scripting

Generate, edit and review Python, C#, VB.NET and IronPython script components. The AI sees inputs, outputs and errors.

docs ↗
script_generate.ghc# · python
AIScriptGen and AIScriptReview components on the Grasshopper canvasscroll zoom · drag pan
screenshot slot
/03

Knowledge & files

McNeel and Ladybug forums, any Discourse community, web pages, PDFs and Office files — ingested as clean Markdown context for the AI.

docs ↗
knowledge.ghforum+files→md
McNeel forum search feeding a Discourse post summarizer componentscroll zoom · drag pan
screenshot slot
/04

Media & Vision

Image generation, image description, canvas and viewport screenshots, audio transcription and speech — all as Grasshopper data.

docs ↗
media.ghimg→text
AIText2Img, Canvas2Img and ImgView components on the Grasshopper canvasscroll zoom · drag pan
screenshot slot
Highlighted — Canvas AssistantFIG.01

Chat with an AI
inside your canvas.

Click the button in the corner of the Grasshopper canvas and talk to an assistant that actually sees your work — components, wires, errors — and can build on it when you ask.

Every edit is staged and reviewable. Multi-document aware. Undo-safe.

reads the canvasedits with reviewundo-safemulti-document
docs ↗
GRASSHOPPER — tower_facade.gh
ParamsMathsSetsVectorCurveSmartHopperMesh
Talk to SmartHopper, your canvas assistant that can see what you are working onscroll zoom · drag pan
screenshot slot — real GH capture here
Highlighted — GhJSONFIG.02
facade.ghjson — patchdiff
// staged change — review before apply
"components": [
{ "guid": "a91f…", "name": "Rectangle" },
- { "guid": "4c2e…", "name": "Brep" }
+ { "guid": "7d10…", "name": "Mesh Box" }
],
"wires": [
+ { "from": "a91f…", "to": "7d10…" }
]

GhJSON — your
definition, as text.

GhJSON serializes a Grasshopper definition to a readable plain-text format — diff it in git, merge branches, apply AI-generated patches, and version-control parametric work like code.

It is the bridge that lets AI read and write your canvas — and SmartHopper is its home: the open spec is created and maintained here.

diff & mergeai patchesversion controlopen spec
guide ↗spec ↗
Highlighted — MCP ServerFIG.03

MCP — let external
agents drive Grasshopper.

SmartHopper exposes its tool set as a local Model Context Protocol server — connect Claude, Codex or any MCP client and let it operate your canvas through the same reviewed tool calls.

Localhost only. Token-protected. Edits staged for review. Every change undoable.

localhost onlytoken-protectededit-gatedundo-safe
docs ↗
CLAUDEexternal agent
localhost
+ token
gate
MCP
SERVER
inside Grasshopper
80+
TOOLS
same tool set
CANVASreviewed edits
Highlighted — Open Source · the commitmentFIG.04
⌥ architects-toolkit / SmartHopper public
licenseLGPL-3.0
stars / forks★ 30+ · ⑂ 4+
source100% public — every tool inspectable
specsghjson-spec · open format
sign-upnone — no account needed
GNU LGPL-3.0 — you may use, study, modify and redistribute. The core will never move behind a paywall or a closed license.

Open source —
always. LGPL.

SmartHopper is free and open source under LGPL-3.0 — and it always will be. The code, the tools, the GhJSON spec: all public, all inspectable.

Fork it, audit it, build on it. Your definitions and your providers stay yours.

lgpl-3.0public repoopen specno sign-up
github repo ↗
Tools map — what the AI can do
drag to rearrange · scroll to zoom · hover for details · click group → subgroup → tool
every tool documented — signatures, params, flagsAI tools reference ↗
Get started

Install in
three steps.

Stable builds ship through the Rhino Package Manager. Experimental builds land on GitHub Releases first.

full install guide ↗
01
Package ManagerRun _PackageManager in Rhino 8, enable the Include prereleases option, and install SmartHopper.
02
Pick a providerOpen Settings → choose one of the AI providers — or run ollama or localai locally, free and private.
03
Talk to your canvasClick the AI button and describe what you want to build.
AI providers — bring your own key, or run local
providerabouttypeauthcostsetupsite
openaithe home of the popular chatgptcloudapi keypaidsetup ↗openai.com ↗
anthropicthe power of claude modelscloudapi keypaidsetup ↗anthropic.com ↗
google geminigoogle multimodal — generous free tiercloudapi keyfree tiersetup ↗ai.google.dev ↗
mistralthe european-based ai labcloudapi keypaidsetup ↗mistral.ai ↗
deepseekopen-weight models at low costcloudapi keypaidsetup ↗deepseek.com ↗
openrouterone key for hundreds of modelscloudapi keyfree + paidsetup ↗openrouter.ai ↗
ollamarun models locally — free & privatelocal—free · privatesetup ↗ollama.com ↗
localaiself-hosted openai-compatible apilocal—free · privatesetup ↗localai.io ↗
Use it for free — pick your path
local

Ollama

The easiest local runner — one installer, any open model. Fully private, no API key, no usage cost.

self-hosted

LocalAI

The open-source, drop-in OpenAI alternative — run it yourself for a fully private, key-free setup.

free key

Google Gemini

Google AI Studio issues a free API key with generous limits — fastest way to try cloud AI.

free tier

OpenRouter

Free-tier open models through OpenRouter — frontier-class models with provider-set rate limits.

FAQ
Is SmartHopper free?
The plugin is completely free and open source (LGPL-3.0), and it always will be. Bring your own API key and you only pay your AI provider for usage — or run local models through Ollama or LocalAI and even that is free.
Which AI providers work?
Eight providers with your own key: OpenAI, Anthropic, Google Gemini, Mistral, DeepSeek and OpenRouter in the cloud, plus Ollama and LocalAI locally. Each provider has a short setup guide.
Does it work on macOS?
Yes — SmartHopper runs on Rhino 8 or newer, on both Windows and macOS. Same plugin, same features.
Is my data sent anywhere?
Only what you ask for, and only where you chose:
  • Your own API key — requests go straight to the provider you picked. Keys are stored encrypted on your system.
  • Local models — with Ollama or LocalAI, your work never leaves your machine.
What is GhJSON?
A friendly, plain-text format for Grasshopper definitions. Once a definition is readable text you can diff changes, keep it in git, and let the AI propose small patches instead of rebuilding from scratch. SmartHopper is its home — we created the format and maintain the open spec.
What is the MCP server?
SmartHopper can run a small local endpoint that speaks the Model Context Protocol — the standard way external AI agents talk to tools. That lets agents like Claude or Codex drive Grasshopper through SmartHopper's toolset. It listens on localhost only, asks for an access token, and won't touch your canvas unless you allow it.
Community

Built for the community —
will you help?

Every great innovation starts with a single contribution. Whether you're reporting bugs, suggesting features, contributing code, sharing example definitions and workflows, or creating tutorials — your voice matters.