SketchUp MCP is a Model Context Protocol server that lets an AI agent read and edit a SketchUp model — creating geometry, moving components, assigning materials and managing tags through plain-language instructions instead of the mouse. There are three ways to get it, and they behave very differently.
The Model Context Protocol (MCP) is an open standard for connecting AI agents to external tools. A SketchUp MCP server exposes SketchUp's modelling operations as tools the agent can call, so an assistant such as Claude can inspect what is in a model and change it.
In practice a SketchUp MCP setup has two halves. A Ruby extension runs inside SketchUp and executes commands against the live model; an MCP server process translates the agent's tool calls into those commands. The agent never writes raw geometry — it calls typed operations, which is why results are more reliable than asking a model to generate Ruby code and hoping it runs.
What differs between implementations is where the model lives while it is being edited. That single difference decides whether the output drops into the project you are working on or arrives as a separate file you have to merge.
Three routes exist as of September 2026. They are not variations of the same thing.
| Open-source MCP servers | Trimble SketchUp Claude Connector | ZirenAI | |
|---|---|---|---|
| Who makes it | Community (several projects) | Trimble, official | ZirenAI |
| Edits your open file | Yes | No — builds in a cloud session | Yes |
| What you get back | Changes in your model | A new .skp to download | Changes in your model |
| Setup | Package runner, JSON config, Python version matching | Sign in with Claude | The app sets it up |
| Needs a Claude subscription | Bring your own | Yes | No — use points in the client |
| What you pay for | Your own AI plan. The server is free | A paid Claude plan | Top up points for actual model usage |
Trimble announced the SketchUp Claude Connector on 28 April 2026. Its published behaviour is to build geometry in a cloud SketchUp session and return a downloadable .skp with a preview thumbnail.
This is the question worth answering before anything else, because it decides how much work lands back on you.
A cloud-based agent builds in a session on a server. You describe a room, it produces a model, and you download a file. If you were starting from nothing, that is fine. If you had a project open — a hundred components, your office tag structure, scenes set up for the next client meeting — you now have two files and a merge to do by hand. Re-scaling, re-tagging and fixing what the import flattened routinely takes longer than the modelling did.
A local agent issues operations inside the SketchUp already running on your machine. The change lands in the model on your screen, Ctrl+Z undoes it like any other edit, and tags, components, scenes and materials are untouched because nothing was ever exported. Open-source MCP servers and ZirenAI both work this way; the difference between them is setup, not architecture.
With the open-source route you install a package runner, edit a JSON configuration file to register the server with your MCP client, copy a Ruby extension into SketchUp's plugins folder, and make sure the Python version on your machine matches what the server expects. It works, and it is free. It is also where most people stop.
Full setup guide → · If the connection fails →
Geometry that is 80% right is often worse than no geometry, because cleaning it up takes longer than modelling it yourself. The work worth handing over is the work where being 80% right is still useful:
It is a poor fit for detailed joinery, topology decisions on curved surfaces, and anything that ends up on a stamped drawing. It moves geometry; it does not have taste.
ZirenAI supports SketchUp 2021 and newer on Windows 10 and 11. macOS is in testing. The open-source servers vary — some require SketchUp 2024 or newer, and the Ruby API changes between versions often enough that a server written for one release may not load in another.
The software is: the open-source MCP servers cost nothing to install. The modelling is not — every route needs an AI client or API key, and that is yours to pay for. ZirenAI uses point packs, with points deducted for actual model usage.
Your description of the task and any reference image you attach are sent to the model provider. The model file itself stays on your machine — the agent issues operations against it rather than transferring it.
Yes. A PDF, a photo or a CAD export can be attached and used as the basis for the geometry. This is the most-used workflow and the one ZirenAI tunes hardest.
Undo. Every run is a normal SketchUp undo step, and each operation the agent performed is listed so you can see where it went off course.
Yes — Rhino has its own MCP route, including an official server published by McNeel. See Rhino MCP.
Open a real project, give it one of the jobs above, and see whether the result is something you would keep. That is the only test worth running.
Download for Windows