AI for CAD Tools
An engineer's look at Zoo's Text-to-CAD feature: how Zookeeper and KCL turn a prompt into parametric geometry, what it still cannot do, and where it fits next to a PDM-aware AI layer.
·
⏱
9 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Zoo's Text-to-CAD feature, built on its own KCL language and 3D kernel, generates genuinely parametric, editable CAD from natural language rather than a static mesh, which sets it apart from earlier text-to-CAD attempts. It still cannot validate manufacturability, has no view into a team's existing parts or PDM history, and is scoped to single parts rather than assemblies. For net-new geometry, and for engineers who want to keep editing what the model produces, it is a capable tool worth trying on the free plan. For checking whether new geometry is needed at all, an AI layer with access to an organization's own CAD and PDM data, like Leo, answers a different and often earlier question.
Zoo, the CAD platform formerly known as KittyCAD, ships a natural-language design agent called Zookeeper. Type a description of a bracket, a flange, or a housing, and Zookeeper writes KCL, Zoo's own scripting language, and hands back a parametric solid with a feature tree an engineer can open and edit. It is one of the more capable text-to-CAD tools on the market, built on a proprietary 3D kernel rather than bolted onto someone else's CAD file format.
This review covers what Zookeeper actually produces, how the underlying KCL language changes what is possible compared with earlier text-to-CAD attempts, what the tool still cannot do for a production engineering team, and where a name like Adam or Spectral Labs' SGS-1 fits as an alternative. It closes with where a retrieval layer like Leo complements rather than competes with a generative tool like this one. The distinction matters because "text-to-CAD" gets used loosely across very different products, from quick mesh generators aimed at hobbyists to tools that, like this one, target production engineers who need an editable, dimension-accurate model rather than a pretty render.
What Zookeeper Actually Generates
Zookeeper is not a single-shot image generator wearing a CAD skin. Every shape it produces is written out as KCL, Zoo's own parametric scripting language, which means the output is a real feature tree rather than a static mesh. A cylinder with a counterbore, a mounting plate with a bolt pattern, or a bracket with a fillet radius all arrive as editable operations: extrude this sketch, cut this hole pattern, fillet this edge. An engineer can open the result and change a dimension the same way they would in a hand-built model, because underneath, it is one.
The workflow is deliberately incremental. Instead of writing one long paragraph and hoping the model reads your intent correctly, Zoo's own documentation describes the process as a series of structured natural-language commands: describe a base feature, review it, then add the next operation. Zookeeper runs in three modes (Auto, Thoughtful, and Standard) that trade speed for how much reasoning the model applies before committing to geometry, and it can produce parametric variants of the same part on request, which is useful for a quick design-space comparison early in a project. In practice, that incremental loop is closer to pair-programming a part with a junior colleague than to a single magic prompt: you steer each step, catch a misread dimension before it compounds into the next feature, and only move on once the current operation looks right.
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. The search system is smart and CAD-aware.
- Eytan S., R&D Engineer
The KCL Language Underneath It
Most CAD kernels were never designed to be read or written by a language model. Their internal representations are opaque binary structures built for a mouse-driven UI, which is part of why early text-to-CAD tools defaulted to producing dumb meshes instead of features. Zoo's answer was to build KCL and a companion 3D kernel from scratch, specifically so an LLM has a readable, generative target to write into. Every feature created inside Zoo Design Studio, whether typed by a person or generated by Zookeeper, compiles down to the same KCL.
That choice is the real differentiator behind the tool, more than the prompt box itself. It is also where the current rough edge shows up: sketches defined inside a custom KCL function cannot be edited with the point-and-click sketch tools, only at the top level. For a team that wants to keep tweaking a generated design by hand after Zookeeper produces it, that constraint decides how the model needs to be structured from the first prompt, not after the fact, which is also why the strongest generative CAD tools right now still sit inside a broader category worth surveying in the best AI generative mechanical design tools for engineers.
Where It Still Falls Short for Production Engineering
Zoo is explicit that Zookeeper cannot run physical tests or guarantee manufacturability or structural safety on anything it generates. That is a reasonable limit for a generative tool, but it means every part still needs the same DFM and tolerance review a hand-modeled part would get before it goes anywhere near a shop floor, and it is worth reading how text-to-CAD accuracy actually holds up against production tolerances before trusting a generated dimension outright. The platform's own guidance is that it works best on solid-modeled parts and well-understood mechanical families; surface-heavy, organic geometry is still an earlier-stage capability.
The bigger gap for an engineering team is context. Zookeeper reasons over the prompt and the active project. It has no view into your approved vendor list, your existing part library, your PDM revision history, or the calculation an engineer already ran on a similar bracket eighteen months ago. Every part it generates is new geometry, which means every part is also a candidate for a design review to check whether something close enough already exists in inventory. Assemblies, multi-part fits, and cross-component tolerance stacks are also not what the feature is built for, a limitation covered in more depth in why text-to-CAD struggles with assemblies; Zoo's own scope is closer to single-part generation than full assembly synthesis. None of that is a knock on the tool doing what it says it does; it is a reminder that "AI generated the part" and "the part is ready to release" remain two different milestones, separated by the same review gates that applied before any of this software existed.
Pricing and How Teams Actually Access It
Zoo Design Studio's free plan includes a monthly allotment of Zookeeper reasoning time with no traditional CAD license required, which is a low-friction way to try it. Plus and Pro plans raise that included allowance and add access to the Thoughtful and Standard reasoning modes; Enterprise pricing is custom and scoped to deployment and support needs. The included monthly usage resets on the account's monthly anniversary and does not roll over, so a team that leans hard on Zookeeper in one week can hit a wall before the month is out unless pay-as-you-go billing is turned on.
For teams calling the API directly rather than using the app, billing works by the second at the plan's current rate for usage-based endpoints; plain GET requests and anything under the user or org paths are free, and a failed call is never charged. Export is available in STEP and STL, which covers the two formats most downstream tools and manufacturing partners expect, and matters most for a team weighing adding a text-to-CAD API into an existing SolidWorks workflow rather than working inside the Zoo app itself.
When to Reach for Zoo, and What a Retrieval Layer Adds
Zoo's Text-to-CAD is a strong choice when the job is genuinely new geometry: a first-pass bracket, a quick parametric variant to compare in a design review, or a concept that has no prior art in your system to draw from. It sits alongside other named text-to-CAD tools such as Adam and Spectral Labs' SGS-1, each with its own tradeoffs on speed, editability, and how much of the output is truly parametric versus a cleaned-up mesh; a full side-by-side of those tools is worth reading before picking one for a specific job.
Where a tool like Leo fits is upstream of that decision. Before generating anything new, Leo searches an organization's own CAD, PDM, and PLM history for a part that might already solve the problem, which avoids paying reasoning credits and review time on geometry that was sitting in the vault the whole time. That is a different job from Zookeeper's: one produces new parametric CAD from a description, the other finds out whether you needed new CAD at all. Run them in that order on a real project and the two tools stop competing for the same step in the workflow; instead, the search step quietly filters out however much of the backlog turns out to already exist, and Zookeeper is left generating only the parts that are genuinely new.
FAQ
See what's already in your vault
Leo searches your CAD and PDM data before anyone designs something new.
Leo finds parts you already have, cites where a spec or tolerance came from, and works across the PDM and PLM systems your team already uses.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
