I keep seeing people in this thread lamenting that "There is no good way for LLMs to design/generate parametric STEP models yet", while I've spent the past 6 months working on this exact problem.
So, yesterday, I've had Astra build a moving assembly of Babbage's difference engine with it with 400+ parts and real parametric geometry, and you can also download the full STEP model assembly there.
Completely free and open source, you can use your own Claude/Codex to start vibe-coding and generate perfect STEP file right now.
So, no, it's not a matter of time on "when can AI do CAD", we are already here.
Sorry if I sound a little frustrated and a bit too self-promotion-y, but I just really want people who says "I wish this capability exists sometimes in the future" to know that it already does, and you can use it right now, for free.
CAD is a different beast than mesh modeling. For product design you need a parametric model that can have precise dimensions and is constructed mathematically step by step. Mesh modeling is just moving a pile of triangles around in a very imprecise way. Works fine for 3D printing or digital assets, but if you want something that can be manufactured that's not good enough. And there isn't a good way to go from STL -> Parametric CAD so there's still room for a product here.
" if you want something that can be manufactured that's not good enough"
People design all sort of buildable things in SketchUp and it's just about dragging polygons around.
You _may_ need accurate surface if you run your model to CAM.
But to this day lot of manufacturing is still done via drawings and there the only thing that matters is that can you draw some type of lines and attach clear dimensions to them. So the shapes don't necessarily need to be accurate for all flows since they are more like topological hints to the person reading the drawing rather than accurate replica of the thing to manufacture.
The only way to go from STL to BREP is through decompilation style recovery/reconstruction, because of the semantic information loss, so it's a task that can only be done via LLMs or human CAD users. Essentially, they pretty much have to remake the model from scratch using the STL as a reference to make the CAD. There really isn't a good way to automate this without machine learning/LLMs.
CATIA has had mesh import with canonical feature detection, and automatic replacement with parametric part features, for at least 25 years. It's in the "Quick Surface Reconstruction" and "Basic Surface Recognition" tools. They even work sometimes.
It's just a matter of time. Astra was trained on a new 3D modelling dataset. They will surely train it on CAD apps (if they have not already done so). I don't hold out much hope for competitors hoping to get in front of that train.
I suppose you might be able to force a Blender agent to do parametric modeling via geometry nodes. It would be a horrible, kludgy, slow workflow for a human, but the AI won't complain.
I was able to have Astra accurately solid model with plasticity (NOT import from Blender) and presumably it would be trivial with autocad, but I haven’t tried it yet.
I had a similar experience, but it worked just well enough that I wanted to try my own similar experiment, but using a functional API in Elixir and giving the agent some introspection tools. I'm very much a novice to this space and learning as I go, but I've been able to make some interesting projects with the library so far, which is all I really want to do: https://smith.hexdocs.pm
If you want to build by hand, there is a Kino renderer to build your projects interactively in Livebook, which is actually surprisingly helpful for learning.
Exactly. I just give Astra an image such as furniture, and it creates an 3D reproduction in Blender. You don’t even need computer use. Codex and Blender’s scripting engine is enough.
I've had decent success getting Claude Opus to design models using the python cadquery library, and then I can import the step file to fusion 360 to make any changes I want afterwards.
I'm so excited about this sort of stuff. ForgeCAD [1] for me was the first mind-blowing application of this a few months ago, where I'm able to just have claude (or codex) write JS that creates 3D Models that I can then print in my 3D printer.
- Built a ring for my compost bin that can properly hold down the plastic bag. This was originally a multi week project because I was also learning Fusion 360 in small 2 hour chunks over many weekends, then with ForgeCAD I had a good prototype in 60 mins with multiple iterations.
- I also built a cover for my overheating Logitech Circle doorbell, snapped a photo, gave it to claude, and claude found all the dimensions on Logitech's website, and then we iterated with 3-4 prints until we had it just right.
This is a whole new set of workflows that are unleashed with agents, even for hobbyists.
From my experience trying to build parametric CAD modeling using FluidCAD [1] using its MCP, the biggest breakthrough is ASTRA. It seems to work well actually with any CAD if it has ways to verify its output. No otjer models was able to match its accuracy even Fable.
They're saying it as if it were contrary to what's being claimed. The tagline says "for architecture and product design." The supported file formats are not video game formats.
I love to see more of these tools build on rhino's kernel. Arguably the best modelling (parametric, manual, or SDK) ergonomics when it comes to, from my experience, early stage aec workflows. A similar tool launched recently [0] that i found interesting. I'm also biased as it's my first and only true love when it comes to "coding things" for the built environment.
My take was that designing buildings was always the cheap part, and not core to what moves the needle in AEC (architecture, engineering, construction)
Minecraft style designs work but complexed geometry yields presentation quality in a triangular hell, which is and has been the problem with all Ai CAD since day 1. What Ai is good at in CAD is finding applicable formulas but anything functioning in a civil space is a nightmare on wax. There's not even enough correct Ai in sensory to make it reliable for self-driving cars and now they want to make the car itself with Ai? I guess in for a penny...
GPT-6 Astra with computer use has gotten some of architects and 3d modelers wondering what their future is. Astra does such a good work with SketchUp 2026 and computer use, here is video, but it is well worth it because the author actually knows his stuff:
The video game applications are truly pretty cool ... though I don't know who would admit to using such a tool in the current environment.
Some of the other things like the `Residential`, `Structure and Complex Shape`, and `Large Scale` seem like they'd need review by engineers to be useful in real world scenarios? Do architects just sketch whatever looks good as a first pass?
SketchUp does not really have that either if you are expecting something else than triangles. Since sketchup has polygons you could argue they are brep but even then only accidentally.
Architecture works with a bit different ruleset than mechanical cad.
Actually this appears to output 3DM, which is the CAD format Rhinoceros 3D uses, and is based on openNURBs. The typically tiny Claude (?) text (such as I can read it) makes it clear it's not just a mesh format, mentions faces and edges.
On that basis it is a step up from the usual AI CAD silliness.
These products are still in their early stages, but it is clear that the AI wave is reaching 3D modeling as well.
It will impact mechanical and civil engineering just as it has impacted software development.
3d modeling for use in software (games or similar) imo is fine, but I would love to not see this stuff in civil engineering. People without AI are unable to design buildings that don't collapse (e.g. recent NYC apartment highrise), adding AI into the mix will just make things worse.
Why do you presume it will make it worse instead of better? Humans, as you point out, could not manage to competently review buildings that don't collapse. Why would adding an additional level of checks (human or AI) do worse than not adding one?
I don't think so. As a mechanical engineer, CAD is only half of the design process. The other half is actually building and assembling your design and testing/debugging what you've created. You still need the mechanical knowledge and intuition to take something in the physical world, figure out why it doesn't work, and apply that to the CAD design to improve it. It would be incredible to have AI help me model faster, but there's no validation loop like there is with software, so you can't have an agent run wild and endlessly iterate.
I'm not a mechanical engineer but I would be surprised if it was even 30% of the process. Materials, manufacturability, durability, fitting into other requirements. Some of those things can be part of the CAD design, but there's so much more.
None of the things you list in the "other half" are out of reach for AI.
Also: disagree about the lack of validation loop:
a) AI can actually build one for you in many cases
b) if you give agents an actual validation toolkit, you'll be surprised what AI can do
Of course, if you expect to prompt "give me a NURBS model of a complete, working v8 motor" and get a result, then yeah, sure, we're not there yet.
I've actually built a bunch of stuff with a combo of astra, blender, openscad and tiny bits of AI written helpers ... giving the AI the proper "hands" to work with the models is key.
One particularly useful tool to provide the AI with is to be able to "look" at the model as it iterates on it (measure dimensions, shoot arbitrary rays at it that return exact intersections and normals, see if the object obeys engineering constraints, etc ...).
Works pretty well for many, many things and saves a TON of time.
AI cannot physically build anything in the real world, what are you referring to? That's a huge part of mechanical engineering.
And you can give it simulation tools, but those are far far away from accounting for everything in real-world physics. You can simulate very tightly constrained problems, but you can't simulate "assemble this actuator and test it", there are so many more variables in the real world that are not easy to account for in a simulation.
If you aren't an engineer that works on physical products, it's hard to articulate just how far from reality CAD and simulations are.
This is how LLMs have been since the beginning. They can be really okay at the most surface level information, but as soon as you dig any bit beyond that, it all falls apart. Inquire any chatbot about something that you're an expert in and it becomes quickly apparent.
I'm here for using LLMs to take some of the tedium out of product design/mechanical engineering, but I just haven't seen a good proof of concept yet. It always seems to be some variation of "Design a sphere. Oh wow, this is incredible!"
Indeed — I try to stay away from website design critique here and it's against a rule, but that dot field makes it actually impossible for me to read, especially given the low text contrast and small fonts on the body text.
I love the web design though! Formas.ai has been doing great work in the visualization space, and it’s great watching them tackle 3D via the very apt name Cartesian. Defo support what they are doing. Maybe I can finally sleep better too!
I've just started working on 3D modeling software of my own, though this looks to be far ahead. Mostly focused on 3D printing - creating printable models is pretty expensive and tedious at the moment, and fully native.
I keep seeing people in this thread lamenting that "There is no good way for LLMs to design/generate parametric STEP models yet", while I've spent the past 6 months working on this exact problem.
So, yesterday, I've had Astra build a moving assembly of Babbage's difference engine with it with 400+ parts and real parametric geometry, and you can also download the full STEP model assembly there.
https://aetheris-difference-engine-showcase.yuechenli.worker...
And here's a STEP file that I've had GPT reconstruct of a popular phone with exact parametric dimensions just from the 2D drawing, took like an hour:
https://github.com/yuechen-li-dev/Aetheris/blob/master/sampl...
Completely free and open source, you can use your own Claude/Codex to start vibe-coding and generate perfect STEP file right now.
So, no, it's not a matter of time on "when can AI do CAD", we are already here.
Sorry if I sound a little frustrated and a bit too self-promotion-y, but I just really want people who says "I wish this capability exists sometimes in the future" to know that it already does, and you can use it right now, for free.
I've been using Astra to just create things in Blender. It's great.
Kill the scroll jacking. It's a sign you're on Windows.
CAD is a different beast than mesh modeling. For product design you need a parametric model that can have precise dimensions and is constructed mathematically step by step. Mesh modeling is just moving a pile of triangles around in a very imprecise way. Works fine for 3D printing or digital assets, but if you want something that can be manufactured that's not good enough. And there isn't a good way to go from STL -> Parametric CAD so there's still room for a product here.
" if you want something that can be manufactured that's not good enough"
People design all sort of buildable things in SketchUp and it's just about dragging polygons around.
You _may_ need accurate surface if you run your model to CAM.
But to this day lot of manufacturing is still done via drawings and there the only thing that matters is that can you draw some type of lines and attach clear dimensions to them. So the shapes don't necessarily need to be accurate for all flows since they are more like topological hints to the person reading the drawing rather than accurate replica of the thing to manufacture.
The only way to go from STL to BREP is through decompilation style recovery/reconstruction, because of the semantic information loss, so it's a task that can only be done via LLMs or human CAD users. Essentially, they pretty much have to remake the model from scratch using the STL as a reference to make the CAD. There really isn't a good way to automate this without machine learning/LLMs.
CATIA has had mesh import with canonical feature detection, and automatic replacement with parametric part features, for at least 25 years. It's in the "Quick Surface Reconstruction" and "Basic Surface Recognition" tools. They even work sometimes.
https://3dswym.3dexperience.3ds.com/wiki/catia-user-communit...
https://resources.rand3d.com/insights-from-within/scanned-da...
Unfortunately the P3 license costs like $15k per year, and it is f'in ugly clunky software.
It's just a matter of time. Astra was trained on a new 3D modelling dataset. They will surely train it on CAD apps (if they have not already done so). I don't hold out much hope for competitors hoping to get in front of that train.
I suppose you might be able to force a Blender agent to do parametric modeling via geometry nodes. It would be a horrible, kludgy, slow workflow for a human, but the AI won't complain.
I was able to have Astra accurately solid model with plasticity (NOT import from Blender) and presumably it would be trivial with autocad, but I haven’t tried it yet.
Astra can use build123d, too. You'll get a STEP out of that.
I've tried Sol on build123d and it was completely unusable, I should take another stab with Astra though.
I had a similar experience, but it worked just well enough that I wanted to try my own similar experiment, but using a functional API in Elixir and giving the agent some introspection tools. I'm very much a novice to this space and learning as I go, but I've been able to make some interesting projects with the library so far, which is all I really want to do: https://smith.hexdocs.pm
If you want to build by hand, there is a Kino renderer to build your projects interactively in Livebook, which is actually surprisingly helpful for learning.
> there isn't a good way to go from STL -> Parametric CAD
yet.
as a matter of fact, you should try what AI can do on that problem, you'd be surprised.
I have tried, I really wish it would be better because CAD is a slog and I want it to be automated. It's pretty bad at real world tasks.
I gave some very bad sketch made in jspaint to Astra, plus some text instructions, and I got back some clean STL file and a beautiful 3D render.
It even made a second STL file with the same model rotated for 3D printing.
I remember all those hours on SolidWorks many years ago, and now you can just speak and present bad drawings to the computer, and it does it okay.
Exactly. I just give Astra an image such as furniture, and it creates an 3D reproduction in Blender. You don’t even need computer use. Codex and Blender’s scripting engine is enough.
Can you do this with the 20 dollar plan?
I have an example here. Took 3 hours with Pro Plan.
https://simulationcorner.net/chair/chair-evolution.gif
Yes, but it eats your 5-hour quota really quick.
Are you hooking it up to blender via some MCP or are you just creating files and opening in blender?
I've had decent success getting Claude Opus to design models using the python cadquery library, and then I can import the step file to fusion 360 to make any changes I want afterwards.
I'm so excited about this sort of stuff. ForgeCAD [1] for me was the first mind-blowing application of this a few months ago, where I'm able to just have claude (or codex) write JS that creates 3D Models that I can then print in my 3D printer.
- Built a ring for my compost bin that can properly hold down the plastic bag. This was originally a multi week project because I was also learning Fusion 360 in small 2 hour chunks over many weekends, then with ForgeCAD I had a good prototype in 60 mins with multiple iterations.
- I also built a cover for my overheating Logitech Circle doorbell, snapped a photo, gave it to claude, and claude found all the dimensions on Logitech's website, and then we iterated with 3-4 prints until we had it just right.
This is a whole new set of workflows that are unleashed with agents, even for hobbyists.
1: https://forgecad.io
From my experience trying to build parametric CAD modeling using FluidCAD [1] using its MCP, the biggest breakthrough is ASTRA. It seems to work well actually with any CAD if it has ways to verify its output. No otjer models was able to match its accuracy even Fable.
[1] https://fluidcad.io
Looking at a bunch of these: I don't think they are even manufacturable. This is something you would use in a video game.
You are right to push back on that.
> This is something you would use in a video game.
You say it like it's a bad thing.
It is when the tool's tagline is literally:
> an AI 3D modeling tool for architecture and product design.
They're saying it as if it were contrary to what's being claimed. The tagline says "for architecture and product design." The supported file formats are not video game formats.
I love to see more of these tools build on rhino's kernel. Arguably the best modelling (parametric, manual, or SDK) ergonomics when it comes to, from my experience, early stage aec workflows. A similar tool launched recently [0] that i found interesting. I'm also biased as it's my first and only true love when it comes to "coding things" for the built environment.
My take was that designing buildings was always the cheap part, and not core to what moves the needle in AEC (architecture, engineering, construction)
[0] https://illoca.com/ , not affiliated
Minecraft style designs work but complexed geometry yields presentation quality in a triangular hell, which is and has been the problem with all Ai CAD since day 1. What Ai is good at in CAD is finding applicable formulas but anything functioning in a civil space is a nightmare on wax. There's not even enough correct Ai in sensory to make it reliable for self-driving cars and now they want to make the car itself with Ai? I guess in for a penny...
GPT-6 Astra with computer use has gotten some of architects and 3d modelers wondering what their future is. Astra does such a good work with SketchUp 2026 and computer use, here is video, but it is well worth it because the author actually knows his stuff:
https://www.youtube.com/watch?v=E-ECbD14g_8
The video game applications are truly pretty cool ... though I don't know who would admit to using such a tool in the current environment.
Some of the other things like the `Residential`, `Structure and Complex Shape`, and `Large Scale` seem like they'd need review by engineers to be useful in real world scenarios? Do architects just sketch whatever looks good as a first pass?
i can't even read the website because you're animating the background, why would i trust you with architecture?
"Architecture and product design" but no STEP/IGES/BRep output, just triangles?
"no STEP/IGES/BRep output"
SketchUp does not really have that either if you are expecting something else than triangles. Since sketchup has polygons you could argue they are brep but even then only accidentally.
Architecture works with a bit different ruleset than mechanical cad.
Actually this appears to output 3DM, which is the CAD format Rhinoceros 3D uses, and is based on openNURBs. The typically tiny Claude (?) text (such as I can read it) makes it clear it's not just a mesh format, mentions faces and edges.
On that basis it is a step up from the usual AI CAD silliness.
These products are still in their early stages, but it is clear that the AI wave is reaching 3D modeling as well. It will impact mechanical and civil engineering just as it has impacted software development.
3d modeling for use in software (games or similar) imo is fine, but I would love to not see this stuff in civil engineering. People without AI are unable to design buildings that don't collapse (e.g. recent NYC apartment highrise), adding AI into the mix will just make things worse.
Why do you presume it will make it worse instead of better? Humans, as you point out, could not manage to competently review buildings that don't collapse. Why would adding an additional level of checks (human or AI) do worse than not adding one?
Ai isn't being used to make competent people better but tries to give supposed expertise to people who don't have skill in that area.
hastily adding AI in into the mix will make things worse
its good to be skeptical, better to spur competition and address any limitations to competition functioning quickly
I don't think so. As a mechanical engineer, CAD is only half of the design process. The other half is actually building and assembling your design and testing/debugging what you've created. You still need the mechanical knowledge and intuition to take something in the physical world, figure out why it doesn't work, and apply that to the CAD design to improve it. It would be incredible to have AI help me model faster, but there's no validation loop like there is with software, so you can't have an agent run wild and endlessly iterate.
> CAD is only half of the design process
I'm not a mechanical engineer but I would be surprised if it was even 30% of the process. Materials, manufacturability, durability, fitting into other requirements. Some of those things can be part of the CAD design, but there's so much more.
Not to say that CAD isn't a big chunk!
Agreed, it's probably 10% of the total product development process, but approximately half if you're just considering "design" tasks.
> As a mechanical engineer, CAD is only half of the design process.
Same thing can be said about animation and games.
None of the things you list in the "other half" are out of reach for AI.
Also: disagree about the lack of validation loop:
Of course, if you expect to prompt "give me a NURBS model of a complete, working v8 motor" and get a result, then yeah, sure, we're not there yet.I've actually built a bunch of stuff with a combo of astra, blender, openscad and tiny bits of AI written helpers ... giving the AI the proper "hands" to work with the models is key.
One particularly useful tool to provide the AI with is to be able to "look" at the model as it iterates on it (measure dimensions, shoot arbitrary rays at it that return exact intersections and normals, see if the object obeys engineering constraints, etc ...).
Works pretty well for many, many things and saves a TON of time.
AI cannot physically build anything in the real world, what are you referring to? That's a huge part of mechanical engineering.
And you can give it simulation tools, but those are far far away from accounting for everything in real-world physics. You can simulate very tightly constrained problems, but you can't simulate "assemble this actuator and test it", there are so many more variables in the real world that are not easy to account for in a simulation.
If you aren't an engineer that works on physical products, it's hard to articulate just how far from reality CAD and simulations are.
This is how LLMs have been since the beginning. They can be really okay at the most surface level information, but as soon as you dig any bit beyond that, it all falls apart. Inquire any chatbot about something that you're an expert in and it becomes quickly apparent.
I'm here for using LLMs to take some of the tedium out of product design/mechanical engineering, but I just haven't seen a good proof of concept yet. It always seems to be some variation of "Design a sphere. Oh wow, this is incredible!"
Considering the ability of men to aim in urinals, Cornucopia is really one of the worst idea.
Is the site owners are here: that background and scroll effect is terrible.
Indeed — I try to stay away from website design critique here and it's against a rule, but that dot field makes it actually impossible for me to read, especially given the low text contrast and small fonts on the body text.
Came to the comments to say something similar.
I usually kind of roll my eyes when people complain about visual effects on a website but I wholeheartedly agree. It makes it very difficult to read
adding representation to the silent non-bothered audience, I like the scroll effect here and don't find it terrible
The moving dotted background makes it impossible to read anything..
My thought exactly. What a terrible background!
oh my... for me its sure way to induce motion sickness...
I love the web design though! Formas.ai has been doing great work in the visualization space, and it’s great watching them tackle 3D via the very apt name Cartesian. Defo support what they are doing. Maybe I can finally sleep better too!
I've just started working on 3D modeling software of my own, though this looks to be far ahead. Mostly focused on 3D printing - creating printable models is pretty expensive and tedious at the moment, and fully native.
[1] https://tomasimodeler.com (pardon the AI-slop, it's just a placeholder)
"What you keep is never redrawn. What clashes is said out loud."