AI for CAD Tools
Hands-on comparison of Zoo, Adam, and Spectral Labs SGS-1 text-to-CAD tools. What each actually outputs, where they fall short, and what engineers need instead.
·
⏱
9
Michelle Ben-David
Mechanical engineer and technical writer specializing in CAD, manufacturing processes, and engineering productivity.
BOTTOM LINE
Zoo, Adam, and Spectral Labs SGS-1 each represent a different bet on how text-to-CAD should work. Zoo generates mesh geometry that works for concept exploration. Adam produces parametric code that works for script-based workflows. SGS-1 is pushing toward native BREP output that could eventually change the game for production engineering.
But all three share the same blind spot: they generate new parts without checking whether suitable parts already exist in your organization. For most engineering teams, the bigger productivity win is not generating new geometry faster. It is finding and reusing the validated geometry you already have. Leo AI's geometry-aware search across your existing PDM vault delivers exactly that, backed by SOC-2 certification and trained on over a million pages of engineering standards.
I have been testing text-to-CAD tools since they started showing up at trade shows, and 2026 is the first year where multiple players have something real to compare. Not demos. Not renders. Actual geometry you can download and open in a CAD environment.
That said, "real" and "production-ready" are not the same thing. The three tools getting the most traction right now are Zoo (formerly KittyCAD), Adam by Datagrok, and Spectral Labs' SGS-1 model. Each takes a fundamentally different approach to the same problem: turning a text description into 3D geometry. I spent several weeks testing all three with the same set of prompts, ranging from simple brackets to multi-feature housings with specific tolerances.
Here is what I found, what surprised me, and why the most useful AI capability for part creation might not be generation at all.
Zoo has been the most visible player in the text-to-CAD space. Their approach is built around an open-source geometry kernel and a machine learning model trained specifically on mechanical part geometry. You type a prompt, the model generates a 3D shape, and you download it.
For basic geometry, Zoo does a decent job. I prompted it for a simple L-bracket with two mounting holes, and the output was recognizable and dimensionally reasonable. A flanged bearing housing came out looking like it understood what a flange is supposed to do. The geometry is clean enough for concept visualization or 3D printing a prototype.
Where Zoo struggles is anything beyond single-body simplicity. Multi-feature parts with specific dimensional relationships get inconsistent fast. I asked for a sensor enclosure with cable routing channels and snap-fit features, and the result was a shape that vaguely resembled an enclosure but had walls intersecting at odd angles and snap features that would never actually function. The output is mesh-based, meaning you get an STL file, not a parametric solid you can edit in SolidWorks or Creo.
For concept-stage exploration and visual communication, Zoo is the furthest along. For production engineering, the mesh output and lack of parametric control remain dealbreakers.
IN PRACTICE
The geometry search has been invaluable - helping me find standard parts instead of designing new ones, saving a huge amount of time and effort.
eytan s., R&D Engineer
Adam takes a completely different path. Instead of generating mesh geometry directly, it uses an LLM to write CadQuery or OpenSCAD scripts from text prompts. The output is code that, when executed, produces parametric geometry.
This is clever for a few reasons. The resulting parts are parametric by definition because they are defined by code. You can change a dimension by editing a variable. You can generate families of variants by modifying parameters in a loop. And because the geometry is code-defined, you can inspect exactly what the tool did and why.
The downsides are significant. CadQuery and OpenSCAD are powerful but limited compared to what a full CAD environment can do. Complex surfacing, organic shapes, multi-body assemblies with mates and constraints... these either do not work at all or require so much manual scripting that the AI assist becomes marginal. I tested Adam with a request for a gearbox housing with integrated bearing seats, and the code it generated would not even compile because it tried to use Boolean operations on bodies that did not intersect correctly.
Adam is genuinely useful for engineers who already work in code-defined CAD workflows. If your team uses CadQuery for parametric automation, Adam accelerates that. For everyone else, the learning curve of debugging generated code is steep.
Spectral Labs entered the space with SGS-1, a model that attempts to generate boundary representation (BREP) geometry rather than meshes or scripts. If it works as advertised, this is a big deal because BREP is how real CAD systems represent parts internally. A BREP output means faces, edges, and surfaces that a CAD kernel can actually work with.
In practice, SGS-1 is still early. The model can produce BREP-like output for simple parts, and the resulting STEP files do open in SolidWorks and Fusion 360 without errors for basic shapes. That alone is an achievement nobody else has matched consistently. But the geometry quality drops off sharply with complexity. Fillets do not maintain tangency. Surfaces that should be planar come through with slight warps that cause downstream issues in assemblies.
I tested it with ten different prompts ranging from a flat plate to a multi-pocket machined housing. The flat plate was perfect. The machined housing was unusable. Everything in between fell on a spectrum that correlated directly with geometric complexity.
SGS-1 is worth watching. If they solve the consistency problem for mid-complexity parts, they will have the most production-relevant text-to-CAD output on the market. Right now, it is a promising research prototype.
Every text-to-CAD tool I tested shares the same fundamental blind spot: they treat part creation as an isolated act. You type a description. You get a shape. That shape exists in a vacuum with no awareness of what already exists in your organization.
This matters because most engineering organizations already have thousands of validated parts sitting in their PDM vaults. When an engineer needs a mounting bracket, the right answer most of the time is not "generate a new one from a text prompt." The right answer is "find the one your colleague designed two years ago that already went through testing, has an approved drawing, and has a known supplier."
Studies consistently show that 60 to 80 percent of new parts designed by engineering teams are functionally similar to parts that already exist internally. The duplication happens not because engineers are lazy, but because searching PDM systems is so painful that designing from scratch feels faster than hunting through poorly tagged file structures.
None of the text-to-CAD tools I tested have any integration with PDM or PLM systems. None of them check whether a suitable part already exists before generating a new one. That is a massive missed opportunity.
Leo AI does not generate new CAD geometry from text. It does something that solves a bigger problem: it searches your existing PDM vault using natural language and geometry-aware matching to find parts that already fit your requirements.
Describe what you need in plain language, and Leo searches across your team's entire design history. It understands actual CAD geometry, not just filenames and metadata. You can search text-to-text, text-to-CAD, or upload a model and search CAD-to-CAD for similar existing parts. Leo offers integrations with leading PDM and PLM platforms including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, and Arena PLM.
The parts Leo finds are already parametric. Already editable. Already validated in production. They come with revision history, associated drawings, and manufacturing data. No mesh cleanup. No manual rebuild. No hoping the AI got the wall thickness right.
For the times when you genuinely need something new, Leo's engineering Q&A provides standards-referenced guidance and calculations with visible logic, so your new design starts with a solid technical foundation instead of a black-box shape.
FAQ
Find Parts Before You Generate Them
Leo searches your vault using geometry-aware AI matching.
Leo AI connects to your PDM and PLM systems to find validated parts by shape and description. Engineering Q&A with cited sources, calculations with visible logic. SOC-2 certified, built for mechanical engineers.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
Find Parts Before You Generate Them
Leo searches your vault using geometry-aware AI matching.
Leo AI connects to your PDM and PLM systems to find validated parts by shape and description. Engineering Q&A with cited sources, calculations with visible logic. SOC-2 certified, built for mechanical engineers.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams