We made an AI CAD tool with FreeCAD, would love feedback

acgacg
in General

Hi all - we made a CAD tool that lets you create and edit models by chatting with an AI agent.

This video shows us building a pavilion house with a spiral staircase.

The idea is that instead of doing everything through point-and-click operations, you can describe geometry and iterate in conversation while the model updates. It’s built with FreeCAD and we support export into most formats.

We’re still early and we’re posting here because we’re hoping to get feedback from people who work with CAD.

You can try it here. There’s a free tier and no credit card is required. We’ll be around all day and would be very happy to answer questions about how it works!

Rick & Alex

falken10vdl

Comments

  • doiadoia
    edited April 16

    So, I post along as I try.
    First of you let me enter my prompt on your start page. I describe a small table with a steelframe and other parameters, takes me about 2 minutes. On enter you ask for creating an account. I create the account, receive the confirmation email and confirm. It redirects me to the start page, the prompt is empty.
    Now you would have lost me, if my curiosity wasn‘t too big.

    acg
  • acgacg

    Hi Doia - thanks for the good feedback, we'll fix that asap!

  • doiadoia
    edited April 16

    Here is the prompt:

    A small table 50x50cm with a steel frame from 20x20mm profiles and a thin sheet metal top plate. Color black, height 110cm.

    The model ask for some clarification:

    Before I build it, two quick checks:

    Do you want a simple square standing table with 500 x 500 mm top and overall height 1100 mm?
    Is the top plate thickness of 2 mm acceptable?
    Should the frame be four straight legs with a perimeter support just under the top, all in 20 x 20 mm steel tube, matte black?
    If that matches your intent, I’ll use that as the design brief and build the first clean model.

    I answer all yes and let the model work.
    (I see no possibility of sharing or exporting the chat. That is sad)

    acg
  • acgacg
    edited April 16

    "(I see no possibility of sharing or exporting the chat. That is sad)" - thats good feedback too, thank you!

  • doiadoia
    edited April 17

    The model works and comes back with an answer:

    Created

    Table_Top_Plate with 500 x 500 x 2 mm sheet top
    Table_Leg_1–Table_Leg_4 as 20 x 20 mm legs
    Frame_Rail_North, Frame_Rail_South, Frame_Rail_West, Frame_Rail_East as the upper steel frame
    Build script saved at workspace/build_table_phase1.py:1
    Brief and log saved at workspace/plan.md:1
    Verified

    Table_Top_Plate measures 500 x 500 x 2 mm at z=1098..1100, so overall height is 1100 mm
    Each leg measures 20 x 20 x 1098 mm, starting at floor level
    Upper frame rails are 20 mm square and sit directly below the top plate
    check_solids() returned no invalid geometry
    Notes

    The black finish is set in the script when a view object is available
    Full image-based 3D QA was not available from the local runtime path, but geometric verification passed cleanly
    If you want, I can do one refinement pass next: add foot pads, a lower stretcher ring, or make the frame more like hollow tube members visually.

    I answer with „ Add foot pads. Can you show me a picture of the model?“ I don’t know what the model build, see no way of verified the result. That’s odd. Maybe because I am on a phone. Will load the desktop page.

  • doiadoia

    After loading the desktop website I can see an empty 3D viewport. I press the Download button, nothing happens. (Error: no geometry to share) The model say it has put screenshots into a folder. I don’t see any folder. You lost me.

    Here are some pictures from the end of the chat:


  • doiadoia

    I will try the desktop version later but for now I have the feeling it doesn’t work. Test end.

  • acgacg

    Hi Doia - we've investigated and identified the issue. Apologies! Working on a fix and I'll update you here.

    brunopostle
  • doiadoia

    Here are the pictures from the Ui in a desktop web view for the interested. Also one of the "rendered" pictures. No geometry was created, pictures are all cut off. Only some python files, which obviously won't work stand-alone (missing import statements), so seem to be intended only for the closed source studio environment. My first impression - very underwhelming.
    I also tried building a simple room with a shop facade, egress doors and so on, but quickly gave up. Prompts are eating token like candy, you probably won't come very far trying to create some meaningful work.

    This is a prototype at best. Looking at the company website, which will not really tell me who these guys are, they started some years back building an AI meeting notes service, expanded it to emails and are now trying to find other areas where they can slap AI onto and monetize. Thanks, I pass.



  • acgacg
    edited April 16

    Thanks for the candid feedback on where the tool falls short, Doia. We're working on improvements.

  • doiadoia
    edited April 17

    Follow up: After some digging here are some infos I found searching the web.
    Cogram is a YC company founded by 2 Germans, who made their PhDs in Physics at Oxford and are now based in Berlin -> hard to find on the company website. Company founded in 2021, startet with AI assisted meeting notes service and expanded to emails and bidding proposals. Customer base primarily big architecture firms in the USA and Western hemisphere. Now they need to pivot on their product, because the service became a commodity, offered by almost any other platform in the AEC arena.

    @acg Here are some basic questions for you which I could not find a sufficient answer in the FAQs. Please keep in mind, you posted on the OSArch forum, a kind of audience who not only are technically versed but also want to know how their tools work and how to improve them. And who hate platform lock-in and VC bullshit. But we love open systems. I know, seems like an uphill battle for you.

    • What is the base model your product works with? ChatGPT, Claude, some other?
    • How, in general, looks your system prompts. What do you do on top of the model?
    • How can I influence the process of the tool? How can I customise the tool?
    • How open is the result I produce with your service? And I don't mean some halfhearted export format, but my actual data.
    • What happens if I decide to stop using your service? Am I able to open my complete data and read it, even if I have to write a script myself? Maybe to transfer it into some other platform or tool?
    • Where and how is your data format documented?

    Oddly, you didn't post in the FreeCad forum. Why?

  • doiadoia

    Follow up: After some digging here are some infos I found searching the web.
    Cogram is a YC company founded by 2 Germans, who made their PhDs in Physics at Oxford and are now based in Berlin -> hard to find on the company website. Company founded in 2021, startet with AI assisted meeting notes service and expanded to emails and bidding proposals. Customer base primarily big architecture firms in the USA and Western hemisphere. Not they need to pivot on their product, because their service is now a commodity, offered by almost any other platform.

    @acg Here are some basic questions for you which I could not find a sufficient answer in the FAQs. Please keep in mind, you posted on the OSArch forum, a kind of audience who not only are technically versed but also want to know how their tools work and how to improve them. And who hate platform lock-in and VC bullshit. But we love open systems.

    • What is the base model your product works with? ChatGPT, Claude, some other?
    • How, in general, look your system prompts. What do you do on top of the model?
    • How can I influence the process of the tool? How can I customiseb the tool?
    • How open is the result I produce with your service? And I don't mean some halfhearted export format, but my actual data.
    • What happens if I decide to stop using your service. Am I able to open my complete data and read it, even if I have to write a script myself. Maybe to transfer it into some other platform?
    • Where and how is your data format documented?

    Oddly, you didn't post in the FreeCad forum. Why?

    The forum seems to delete my posts if I try to edit typos. So this is a repost.

    semhustejwalpaJanF
  • semhustejsemhustej

    I would be interested in this

    • under what license are the models created by the tool distributed under?
    Nigel
  • acgacg
    edited April 17

    @semhustej @doia thank you for the candid feedback and for taking the time to test.

    A few answers:

    • Ownership / license of created models: under our terms you own the input prompts, files, and outputs you create (to the extent that you own them under applicable law). That includes 3D models, scripts, renders, and exports. Our terms of service are here: https://studio.cogram.com/terms.
    • Data portability: for now you can import/export .FCStd, IFC, STEP, and other common formats from the Import/Export menu at the top center of the viewer.
    • Our goal is that your model data is not trapped in a closed format.
    • Model stack: under the hood we use language models and the system is not tied to a single provider. At the moment it is OpenAI's GPT 5.4.
    • Customization is still limited today but that's one thing we expect to improve.

    We posted here because we felt OSArch is a technically rigorous audience that will give to-the-point feedback. Thanks again for trying it - we’ve confirmed several issues and are working on fixes. Will update you here when those are resolved.

    semhustej
  • doiadoia
    edited April 18

    Thanks for coming back to us.

    Now, after the criticism some ideas I would find cool if I would use your product:

    • better selection indication with more information about the selected element than just the name
    • some way of precustomizing with e.g. loading skills or setting an account/project wide base prompt (AGENT.md)
    • proper navigation information in the view port - where is positive Z, positive X etc.
    • preview of images/other results not just in a small window - give me the opportunity to view it near full size

    But I will not use your product, at least not in the near future and at this price point without any meaningful value added to the base (and free) app.

    I’d rather look into https://github.com/neka-nat/freecad-mcp or https://spkane.github.io/freecad-addon-robust-mcp-server and fire up my local Llama.cpp instance.

    Good luck with your endeavor.

  • acgacg

    Thank you @doia , those are all good suggestions for improvements.

Sign In or Register to comment.