Hi everyone,

I've been working on a small side project and it's finally ready to share: mcp-bridge-for-moi, a free and open-source MCP server that connects an AI assistant (e.g. Claude Code, Claude Desktop, or any other client that supports the Model Context Protocol) to MoI3D.

You describe what you want, and the agent builds it right in your open MoI document. It looks at the result, fixes its own mistakes, and keeps going. You stay in charge the whole time: select something to point at it, undo a step, or take over whenever you like.

Here's a short video of it in action:
https://youtu.be/OCRmmbYtlfk

And here's the repository with everything you need:
https://github.com/preluceo-tools/mcp-bridge-for-moi

A few things worth knowing
Just to be clear, this is an unofficial project and isn't affiliated with or endorsed by Triple Squid Software.

• The agent never saves your file. Nothing is written to disk unless you save it yourself.
• One Ctrl+Z undoes one whole agent step.
• If you're in the middle of a MoI command, the agent waits until you finish or cancel before it changes anything.
• It needs MoI 4 on Windows and Node.js 22 or newer. MoI 5 and macOS haven't been tested yet. If you have a Mac and feel like trying it, I'd love to hear how it goes.

Setup: a bit of DIY, sorry!

Unfortunately there's no simple installer yet. You'll need to do the technical setup yourself: download the code with git, build it with Node.js, install the bridge script into MoI's startup folder, and add the server to your AI client's configuration. It's all explained step by step in the README. If you've set up an MCP server before, it should feel familiar.

Work in progress

Please treat this as a work in progress. It's tested, but you'll probably still run into bugs. Save your work before you let the agent loose on it, keep backups of anything important, and keep an eye on what it does. The README has a Safety section worth reading before you start.

This is a spare-time project, so I'll fix bugs as my free time allows. If you run into a problem, opening an issue on GitHub is the best way to report it, and it really helps me.

- - - -
▎ v. 0.2.0 is out:

The full change log can be found here: Changelog v 0.2.0
what changed? => the export function is now available in the MCP.
Please update.
A short description on that can be now found on the repo main page.
- - - -
▎ v. 0.3.0 is out:
a new boolean() helper warns when a difference or union missed its target, a script that throws now names the objects it left behind, and get_scene stays within the client's output limit on large scenes. After updating, install the bridge again, because the server refuses the old one.
- - - -
▎ v. 0.4.0 is out
This release only changes the server. The MoI side is unchanged, so you don't need to reinstall the bridge. Pull, rebuild, and restart your AI client.

What's new:

• Better help on MoI objects. moi_factory_help can now describe objects, edges and styles, which MoI's own API doesn't document. It works without a running MoI session.
• Clearer hints when a script goes wrong.
• If a script returns nothing because it has no return, the reply now says so.
• If a script fails on a large model, the error stays short: it lists up to 10 object ids and gives a count for the rest.
• Nine more MoI scripting pitfalls are explained to the agent. These include commit() return values, mesh import, join returning one object per connected group, and isClosed on breps. The whole list is now easier for the agent to read.
• Units. On a document with no unit system, the agent may set millimeters on its own when it's only testing. Otherwise it still asks you first.

Full changelog and download:
Changelog mcp-bridge-for-moi 0.4.0
- - - -
▎ v. 0.5.0 is out

• New tokens command. "mcp-bridge-for-moi tokens" estimates how many tokens each tool definition costs an agent, and the total. It needs no MoI session. Add --json to print the definitions as the server sends them.
• Token use section in the README. It lists the cost tool by tool and explains how to check it in your own client. With every definition loaded up front, the estimate is about 3,900–4,400 tokens. Claude Code's /context counted about 5,700 in my test. Clients that load definitions on demand pay less.
• A size guard. A test fails ifst a set size, so the cost can't creep upunnoticed.
• moi_eval description trimmed.de's 2,048-character cap on tooldescriptions, so its ending was cut off. The MoI scripting traps and helper details now live inmoi_factory_help('traps') and moi_factory_
• Pointer on failure. When a moi_eval fails with moi_error, the reply now points to moi_factory_help('traps').

Release notes and download: v0.5.0 on GitHub'
---
A very special thank you to Michael for MoI 3D.Thank you so much for the incredible tool that has saved my skin over the years.


Happy modelling!
Gabriel

I suppose that you can talk to the AI ? (against write )

When Claude is used does it limited in time ? (in free)

Nice...but to me the most relaxing thing to me is to drawing.

but can it make scripts?
I need Rhino multipipe in Moi :D

Always interesting to watch these things work.

It's also always interesting to watch the tokens cash register Ka-ching its way inexorably upwards

Looks like it cost about 20,000 tokens to get there

Question: I tried to look the cost up for the job and I got INPUT tokens and OUTPUT tokens

sooooooo... are these input or output tokens your program has in the costing? There's a 400% difference.

https://tokencalculator.ai/

Hey Pilou,
I didn't quite understand your first question, but if you mean 'talking' to the agent, then yes, there are tools like 'Whisper Flow' that can help with that.
Unfortunately, you can't use a free Claude account for this – you need a paid tier to access the MCP via Claude Code.

Hey Matadem,
No, scripting capabilities are outside the scope of the MCP server.
However, if you're interested in agentic AI capabilities: I'm sure that Claude (and other AIs, although I'd recommend Claude) could help you develop what you need for MoI.

Hey ado1,
I would highly recommend working with agentic AI on a paid subscription tier, such as a "Claude Pro" account. API requests will cost you a fortune; this is definitely not the way to go about it in my opinion.
As for the MCP tokens, I honestly have no idea — I didn't pay attention to how many it used.
The MCP server is the first part of my plan to automate modelling. My initial idea was to go from a hand-drawn sketch to a NURBS prototype in the shortest possible time. I am currently working on the second part, which will involve taking a hand-drawn or AI-generated image and modelling it in MoI. It's a long shot, but the initial tests are promising.

Yes talk with a microphone! :)
But i don't want use a paid Ai ;)

Don't get me wrong, it looks amazing.

It's those devil in the detail dollar things that people need to know before committing

Your "drawing" route looks like a very interesting route

There are many out of production items from the 20th and 19th century that would benefit from modern methods

edit: and MOI can scale items very precisely, so the initial output only needs accuracy in any scale whatsoever

Pilou,
You can use open-source AI libraries to achieve this.

Ah yes i remember this one! ;)
My Ai process from the some "bad" first image! :) False image of course like Canada dry! :)

ado1,
I understand your point — nowadays, everyone working with AI has to keep an eye on costs.

Your drawing looks intriguing — it might be a nice test drive. Thanks!
Regarding the scale of output, I agree. The main idea behind forcing the question of units in a fresh, unitless setup was to keep the workflow clean for the upcoming skill (e.g. if dimensions are written in a specific unit on the drawing by hand).

Claude understands prety well itsself precive dimensional drawing (video in the header of the GitHub repo show the modeling based on a sample Design Spark drawing:
https://www.rs-online.com/designspark/designspark-mechanical-drawing-add-on-maxing-its-potential

Yes. If you have a few precise measurements for a few easy parts contained in the original drawing then it will give the AI a good leg up.

@ Gabriel

You will not have the sort of same result in using "Windows Voice Access" ?

Pilou,
to my knowledge Windows Voice Access can't print in just any software because it relies entirely on the app's built-in design.
For Voice Access to trigger a print job, the software you are using must do at least one of three things:
- Support standard keyboard shortcuts (like Ctrl + P)
- Have a visible, accessible "Print" button that the AI can read and click
- Be built with standard Windows code structures that allow assistive technologies to "see" and control the menus.

Exactly for that reason, and as you mentioned, to avoid using a paid tier, I started building one on my own with Python and free AI speech-to-text libraries. Unfortunately, it's a closed repository at the moment (it's for my day job, so I can't publish it yet, but hopefully I will be able to in the near future).
If you're interested in AI-driven software development, I can tell you that, as far as I know, there are two major open-source players: "Whisper", published by OpenAI, and "Moonshine". These are the two I've been building my Python project on, and both work pretty well. Note: I was experimenting with 'Faster Whisper', as the name suggests. – it is faster. Here are both repos:

https://github.com/SYSTRAN/faster-whisper

https://github.com/moonshine-ai/moonshine

Cheers!
Gabriel

Ado1,

I was curious to give you a rough idea of the cost per token. Generally, input tokens are the data that you provide to the AI agent (e.g. text or drawings). The output is what the agent gives you back (text answers in the 'chat' — or the 'session', as it is called — and all the output needed to build the 3D model in MoI in this case).

To try out a more complex task, I provided Claude with your sample drawing (it took some thought, but I don't think "high" makes much difference at the current stage of MCP development). I also asked it to research additional sources online for precise dimensional drawings.
It started with my work-in-progress 'Sketch to 3D' skill, found some additional information on the web, and after about 13 minutes it gave me the attached scene.
I completely agree that this is far from the original drawing, with some obvious errors. I created the MCP server mostly for conceptual design, to produce rough 3D shapes as prototypes for different 3D printing projects.
But in terms of tokens: It used around 50,000 tokens on this model, which, on my paid 'Claude Pro' subscription, is around 8% of my usage limit within a five-hour window. This is basically how these plans work: you work, then you wait until the time window unblocks, then you work again.
This means it didn't use up many tokens, which makes me think that many MCP calls to MoI generally don't use up many tokens either.

It is also always interesting to observe the thinking process of an agent. For example, I asked iClaude to summarise all the uncertainties in an MD (Markdown) file, which it saved at the end of the session. I have also attached it here for reference.

Use cases are obviously always a consideration; I'll work on improving my 'sketch-to-3D' skill in the meantime and see what I can come up with.

Cheers,
Gabriel

@ Gabriel for complete that you said

How to set up voice control (workaround):
Since MoI3D relies heavily on keyboard shortcuts (hotkeys) and scripts to launch commands, you can use voice-recognition-based dictation or macro tools on Windows.

Compatible software: Applications such as VoiceMacro, Talon Voice, Dragon, or the native Windows Voice Access tool.

How it works: You map a spoken word to a specific MoI3D keyboard shortcut.

Example: Saying "Line" simulates pressing the "L" key (or whichever shortcut you have configured for the Line tool);

saying "Circle" simulates the "C" key.

Limitations of this method: While voice control is convenient for quickly triggering tools or scripts, it quickly proves inadequate for precision 3D modeling.

Point placement: MoI3D requires precise clicking in 3D space or the entry of exact numerical coordinates. Voice input lacks the precision needed to target vertices or draw smooth curves on the screen.

Therefore, it works well as a supplementary method (for accessing menus or functions) but is difficult to use as a sole means of drawing.

Why does “Claude” need MoI?

It has access to every paper and book and bit of information Michael ever read or studied.

Why dont you just go model in Claude? You can stop, check your work, and resume and save whenever you want!

“DIY install?” Why doesnt Claude take care of that for you?

Claude could make this post for you! Why are you wasting time? Its not the future for you to waste time.

Even “Frenchy” is presenting “commands” in here now…

Let the SPAM begin…….

Hi Burr,

> Why does “Claude” need MoI?

I was fooling with Grok some months back, asking it to produce some dimensioned, scaled architectural models. It quickly turned into a mess, so I began asking "it" about its underlying tools and abilities. What I learned was that none of the then current AIs had integral geometry kernels though there were some efforts underway to communicate with some CAD packages (typically work being done by the CAD vendors).

At that point I shrugged and went back to using MoI as I normally do because for me, drawings without scale and precision are not useful.

I haven't visited Gabriel's github yet, but am a bit curious about how it controls MoI.

--Larry

Hey Larry,
what you're saying is absolutely right, and it's exactly why the MCP approach works differently.

Most commercially available AI models still have no geometry kernel, and mcp-bridge-for-moi doesn't pretend otherwise. It doesn't generate geometry at all. All the geometry is made by MoI's own kernel, in your open document, through MoI's scripting API. The AI's only job is to decide which MoI command to run and with what numbers — the same decisions you make when you type a value into a command's input field.

How it works in practice

Two pieces: a small script that sits in MoI's startup folder and connects out over a local WebSocket, and an MCP server that your AI client (Claude Code, Claude Desktop, anything that speaks MCP) launches. The server talks to the agent on one side and to MoI on the other. It listens on 127.0.0.1 only, with a token regenerated each start.

The agent gets a handful of tools. The important one runs JavaScript inside MoI, which is how it calls MoI's geometry factories — the same 110 commands the UI drives (box, loft, revolve, fillet, boolean, …). 103 of them can be driven from a script; the other 7 need an interactive viewport pick or produce nothing headless, and are documented as such. Every one of the 103 is committed with real inputs and its output checked by an automated test run.

The rest of the tools are: list the scene (id, type, whether an object is a closed solid, bounding size, document units), read your selection, set the selection or the view, render a viewport image so it can look at what it just made, and export to STEP/OBJ/STL etc.

On scale and precision specifically

- Every factory input is a number in document units. A box 3600 × 240 × 2700 mm is that, exactly, because MoI built it.
- If the open document has no unit system, modelling and export are refused with a no_units error until you pick one.
- Exports go through MoI's own translators, so a STEP out of it is a MoI STEP.

So the failure mode you hit with Grok — plausible-looking nonsense with no kernel underneath — isn't really available here. What can go wrong is different and more mundane: the agent picks the wrong command, mis-orders an operation, or misjudges a dimension it inferred rather than was given. You see it happen live in your viewport.

Cheers,
Gabriel