What Is MCP? How to Connect AI to Rhino and Build 3D Models Automatically

MCP (Model Context Protocol) is an open standard that lets AI agents use real software, and it is how you connect AI like Codex to Rhino 3D. Once connected, you describe a building in plain language and the AI builds it in your Rhino file, step by step, as real and editable geometry. This guide explains what MCP is, how to set it up, and shows a live example: Richard Meier's Douglas House, modelled from one prompt.

Quick answers

What is MCP? MCP, the Model Context Protocol, is an open standard that connects AI models to external tools and software. Anthropic introduced it in November 2024, and OpenAI adopted it in March 2025. Think of it as a universal plug: any AI that speaks MCP can use any tool that offers an MCP server.

How do you connect AI with Rhino? Install a Rhino MCP plugin from Rhino's Package Manager, start its server inside Rhino (for example with the mcpstart command), then register that server in your AI agent, such as Codex. The AI can then call Rhino commands directly.

How does AI build a 3D model automatically? You give the agent a brief. It researches references, splits the brief into steps, and calls Rhino tools one by one: terrain, floor plates, walls, glazing, stairs. Rhino runs each call and reports back, so the agent can check and continue.

What is MCP (Model Context Protocol)?

MCP is an open standard for connecting AI systems to the tools, files and software around them. On its own, an AI model can only read and write text. It cannot open your Rhino file, draw a line or move a layer. MCP gives it a common way to see what a program can do and to ask that program to do it.

MCP is often called the "USB-C for AI". Before USB-C, every device needed its own cable. Before MCP, every AI-to-software link needed its own custom integration. With MCP, a tool publishes one MCP server, and any MCP-compatible AI can plug into it.

Key facts:

• Created by: Anthropic, announced on 25 November 2024.
• Adopted by: OpenAI in March 2025, across products including the ChatGPT desktop app, and by other major AI providers.
• Governance: in December 2025 Anthropic donated MCP to the Agentic AI Foundation, a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI.

Why MCP matters for architects

For architects and designers, MCP turns an AI chat into a modelling assistant that works inside the software you already use. Instead of copying scripts from a chat window into Rhino, the AI runs the commands itself and sees the result. The output is not a rendered image: it is real NURBS geometry, on real layers, in your own .3dm file, ready to edit by hand or in Grasshopper.

How MCP works with Rhino

MCP connects four parts in a loop: you, the AI agent, a Rhino MCP server and Rhino itself. The server is the plug in the middle.

You write a brief; Codex breaks it into steps and calls one tool at a time; the MCP server passes each call to Rhino; Rhino runs it and returns the result, so Codex can check its work before the next step.

A Rhino MCP server typically exposes tools like these:

Tool type	What the AI can do
Create and modify objects	Draw points, lines, curves, boxes and other geometry; move, scale, delete
Inspect the document	Read objects, layers and materials in the open file
Manage layers	Create, rename and organise layers
Select objects	Pick objects by type, layer or other filters
Run Python scripts	Execute RhinoScript or Python for anything more complex

How to connect AI with Rhino: step by step

Connecting Codex to Rhino takes about 15 minutes and five steps. The example below uses the open-source RhinoMCP plugin; other Rhino MCP plugins follow the same pattern.

Before you start, you need:

• Rhino 7 or newer (Rhino 8 recommended), on Windows or Mac, updated to the latest service release
• Python 3.10 or newer
• The uv package manager (brew install uv on Mac, or the PowerShell installer on Windows)
• An AI agent that supports MCP: Codex (in the ChatGPT desktop app, the Codex CLI or the IDE extension), Claude Desktop or Cursor

1. Install the Rhino MCP plugin. In Rhino, open Tools → Package Manager, search for rhinomcp and click Install.
2. Start the MCP server inside Rhino. Type mcpstart in Rhino's command line. The server listens for the AI over a local TCP connection.
3. Register the server in Codex. In a terminal, run codex mcp add rhino -- uvx rhinomcp. This writes the server into ~/.codex/config.toml, which the Codex CLI and IDE extension share. You can also add it by hand:

   [mcp_servers.rhino]
   command = "uvx"
   args = ["rhinomcp"]

4. Restart Codex and check the connection. Restart your Codex session so it loads the new server, then ask it to list the layers in your open Rhino file. If it answers with your real layer names, the connection works.
5. Run a first test. Ask for something small, such as "Create a 10 × 10 × 3 m box on a new layer called Test". Watch it appear in the viewport.

Everything runs locally: Codex launches the server from your config, and the server reaches the plugin inside Rhino over a TCP socket.

Using Claude Desktop or Cursor instead? Add the same server to claude_desktop_config.json or .cursor/mcp.json under mcpServers, with "command": "uvx" and "args": ["rhinomcp"]. Run only one AI client against the Rhino server at a time.

Troubleshooting: if the AI says it has no Rhino tools, the server was not loaded: restart the AI client. If it sees the tools but calls fail, Rhino's server is not running: type mcpstart again and keep Rhino open.

How to use AI to build a 3D model automatically: the Douglas House case study

We asked Codex to model Richard Meier's Douglas House (1973), a white, multi-storey house set into a steep slope above Lake Michigan. One prompt produced a complete LOD 200 model in Rhino, built live in front of us. Here is how the AI worked through it.

Step 1: one prompt

The whole brief was a single message:

In Rhino, create a replica of the Douglas House by Richard Meier at approximately LOD 200. You can search for details and specifications online. Show how you build the objects in Rhino in real time.

Asking for a level of detail (LOD 200) and a live build gives the AI a clear target and makes every step visible.

Step 2: the AI researches the references

Before drawing anything, Codex searched the web and viewed photos and drawings of the house. It summarised the design features the model had to capture:

Feature	What the AI found	How it shapes the model
Entrance bridge	Entry is by a bridge from the road at roof level	A bridge object connects the road to the top floor
Hillside levels	The living levels stack down the steep slope	Floor plates step down; the uphill side is set into the ground
Lake-facing glass	A tall glazed volume faces the lake	The west façade becomes full-height glazing with balconies

Step 3: the AI builds it live in Rhino

Codex then called Rhino's tools one by one, through MCP: terrain and slope, floor plates, balconies, structure, glazing, interior and exterior stairs, railings. Each call created real objects on named layers, and Rhino reported back after every step. In our video the build plays at about 2.7× speed.

Codex was also clear about what it did not know. Before building, it stated its working assumptions: an approximate 9 × 18 m organising footprint and 3 m level spacing, "modelling assumptions, not surveyed dimensions". When it finished, the Rhino file held 566 objects.

Step 4: review and iterate

The result is ordinary Rhino geometry: you can orbit it, switch display modes, select and edit any object, or keep prompting. Follow-ups such as "make the balconies 1.5 m deep" or "add mullions every 1.2 m on the lake façade" change the existing model rather than starting over.

Prompt templates for architects

Good prompts give the AI a building, a level of detail, real dimensions and a working method. Copy and adapt these:

Model a reference building

In Rhino, create a replica of [building] by [architect] at approximately LOD 200.
Search for drawings and photos first and list the key features before modelling.
Put each element type on its own layer (terrain, slabs, walls, glazing, stairs).
Build in real time and tell me what you create at each step.

Massing from a brief

Create a massing model for a 60 x 40 m site: 6 storeys at 3.5 m,
a 12 m central courtyard, and a 3 m setback above level 4.
Use metres, put each level on its own layer, and report the gross floor area.

Parametric edit

Add vertical fins to the south facade at 1.2 m spacing, 1.5 m deep.
Then try 0.6 m and 2.4 m spacing on copies so I can compare.

Tips

• Always state units and key dimensions; the AI will otherwise guess.
• Ask for layers by element type, so the model stays easy to edit.
• Ask the AI to research and summarise before modelling, as in the case study.
• Work in steps: massing first, then façade, then details.

Limitations and best practices

AI modelling through MCP is fast for massing, studies and reference models, but it does not replace your judgement. Keep these limits in mind:

• Accuracy depends on the sources. A replica built from online photos is an interpretation at LOD 200, not a survey. Check key dimensions against drawings before you rely on them.
• Complex geometry may need scripting. Some Rhino MCP plugins support only basic primitives directly, and use Python scripts for anything more complex. Ask the AI to show you the script it runs.
• Save often. The AI edits your open file. Save a copy before large changes, and use Rhino's undo if a step goes wrong.
• One client at a time. Connect only one AI client to the Rhino MCP server at once.
• Review what runs on your machine. Install MCP plugins only from sources you trust, and read what a script does before you approve it.

FAQ

What does MCP stand for?

MCP stands for Model Context Protocol. It is an open standard, introduced by Anthropic in November 2024, that connects AI models to external tools, data and software.

Can ChatGPT or Codex control Rhino?

Yes. Codex, OpenAI's coding agent, supports MCP servers. With a Rhino MCP plugin running, Codex can create, edit and inspect objects in an open Rhino file.

Is MCP only for Claude?

No. Anthropic created MCP, but it is an open standard. OpenAI, Google DeepMind and many tool makers support it, and it is now governed under the Linux Foundation.

Do I need to know how to code to use AI in Rhino?

No. Setup needs a few copy-paste commands, and after that you work in plain language. Knowing some Python helps when you want to check or reuse what the AI runs.

Which Rhino version do I need?

The RhinoMCP plugin needs Rhino 7 or newer, on Windows or Mac. Rhino 8 is recommended.

Does the AI produce real geometry or just images?

Real geometry. Every object is native Rhino geometry on layers in your .3dm file, so you can edit it by hand or connect it to Grasshopper.

Can AI work with Grasshopper too?

Yes. MCP connections for Grasshopper let an AI agent build and adjust parametric definitions. Our course covers AI agent plus Grasshopper workflows.

Is it safe to let an AI run commands in Rhino?

It runs only the tools the MCP server exposes, on your machine. Use trusted plugins, save before big changes, and review scripts before they run.

Learn the full AI-to-Rhino workflow

The Douglas House build is one lesson from our AI 3D Modeling course for architects. The course walks through the whole workflow, from setup to parametric design:

1. AI agent 3D modelling: what it can do
2. Set up your workflow: ChatGPT, Codex and Rhino MCP
3. Text to 3D model
4. Image to 3D model
5. Add specialised skills to your AI agent
6. Create your own skill
7. AI agent + Grasshopper

Get the course →

Sources

• Model Context Protocol, Wikipedia: origin, adoption and governance dates
• RhinoMCP setup: requirements, Package Manager install, mcpstart, client config
• Codex CLI MCP setup: codex mcp add and ~/.codex/config.toml