Text-to-CAD Guide 2026: Functional Parts Generative Meshes Cannot Make

Text-to-CAD is a class of generative AI that turns a natural-language prompt directly into a CAD model whose dimensions you can edit afterward. Type “a flange, 60 mm outer diameter, with four M4 bolt holes” and what comes back is not a pile of triangles but engineering data with sketches and a feature history. Even though both are “text to 3D,” the output data structure is fundamentally different from the mesh generators that produce figurines.

That difference is the boundary between decorative prints and functional parts. A generated bracket that does not quite fit the shelf, a bolt that will not pass through, a hole you want to widen by just 2 mm but cannot: everyone who has tried to use mesh generation for functional parts hits the same wall, and it is rooted in data representation, not model quality. Which is why the fix is not “a better prompt” but a different technology.

This article digs into why generated meshes fail at functional parts down to the data-structure level, then walks through Text-to-CAD with Zoo Design Studio, which you can try for free today. For the mesh-generation side of the landscape, see our Text-to-3D guide and the AI 3D generation roadmap.

Where Generated Meshes Stop: Functional Parts

Mesh generators like Meshy and Tripo have become remarkably practical over the past year. Character figures, decorative panels, cosplay props: for prints where looks are the point, the road from prompt to printer is already paved. We covered mesh repair and slicer prep in Making Generated Meshes Print-Ready.

The story changes the moment you ask the same tools for a functional part, meaning a part that touches other parts: a shelf bracket, a camera mount, an enclosure for a power board, a replacement for a cracked OEM piece. What these have in common is that some dimension is constrained by external reality, like a bolt diameter, a rail width, or hole positions on a PCB.

A generated mesh has no concept of that constraint. The hole in the STL is not a “through-hole of a specified diameter” as design intent; it is merely an arrangement of triangles that happens to look like one. Write “hole for an M4 bolt” in the prompt and nothing guarantees the dimensional accuracy of what comes out, and you have few ways to verify it.

The repair stage is even worse. When a printed part does not fit, CAD lets you retype the offending dimension and move on. A mesh records no dimension to retype. Widening a hole by 0.4 mm becomes manual sculpting or a regeneration lottery. This irreversibility is the real reason generated meshes stall at functional parts.

Picture a holder for a power adapter that hangs on a DIN rail: rail width 35 mm, 0.3 mm clearance to the adapter body, 2.5 mm walls. The design lives or dies on those three numbers. If it will not clip onto the rail, the fix is one opening dimension. In CAD that is thirty seconds, and nothing else moves. With mesh generation you rewrite the prompt, regenerate, and start re-measuring output in which parts you never wanted changed are now different. Functional-part design is not about nailing the answer in one shot; it is about iterating. Once iteration is the premise, choosing a data representation with cheap edits matters more than generation quality.

Mesh vs B-Rep: Two Different Things Called a 3D Model

The key to this wall is the two dominant representations of 3D data. One is the mesh, with STL, OBJ, and 3MF as the typical formats: the surface of an object approximated by many small triangles. Curves are recorded not as true curves but as collections of tiny flat facets. Slicers consume meshes directly, and they suit rendering and games.

The other is B-Rep, boundary representation: the object described as mathematically exact faces, edges, and vertices, plus the topology connecting them. A cylindrical face is stored as the equation “cylinder of radius 20 mm,” which is how mechanical CAD like SOLIDWORKS and Fusion represents solids internally. The flagship exchange format is STEP, standardized internationally as ISO 10303.

The last row is the decisive one. B-Rep to mesh is a one-way, trivial conversion; PrusaSlicer and OrcaSlicer can open STEP files precisely because they tessellate them into meshes on import. The reverse, reconstructing “this was supposed to be a 20 mm radius cylinder” from a bag of triangles, is effectively out of reach. Triangles carry no information about where one face ends and another begins, and even if faces could be guessed, constraints like “these two holes must be coaxial” are design meaning that geometry alone cannot recover. What is lost is not shape but intent. So the conclusion is simple: functional-part AI does not need better mesh post-processing. It needs a different pipeline that outputs B-Rep from the start.

What Text-to-CAD Actually Is

Text-to-CAD is the name of that pipeline: natural language in, a B-Rep solid out, usually as a STEP file. Because the output is engineering data rather than a mesh, you can open it in CAD, check dimensions, add fillets, and retype hole diameters, merging straight into the conventional design workflow.

The implementation approach differs from mesh generators too. Instead of learning what shapes look like, it generates the procedure that builds the shape. In the case of Zoo, the machine-learning model emits KCL (KittyCAD Language) code describing the modeling steps, and a geometry engine executes that code to produce the B-Rep. Because the procedure survives as code, the model is inherently parametric: change a value and regenerate.

A quick note on the word parametric. In parametric CAD a model is stored not as final coordinates but as a history of operations: sketch, extrude, drill, fillet. Each operation carries dimensional parameters, and changing one recomputes the whole history. Rewrite a 4.2 mm hole as 5.0 mm and the chamfer that depends on it follows automatically. That is what “an editable 3D model” really means, and it is why a Text-to-CAD result arrives as a reproducible recipe rather than a finished statue. Humans can swap out just the steps they dislike, so the imperfection of generative AI is cheap to fix. Code is also a readable design history, which pairs naturally with version control like Git.

It is not a universal tool. Today Text-to-CAD is strongest on parts whose dimensions are easy to state in words: flanges, brackets, shafts, enclosures. For organic sculptural shapes, mesh generation still wins. The two are not competitors; they split the work by output data type.

Zoo Design Studio: The Practical Way In

Zoo Design Studio is an AI-native CAD platform with Text-to-CAD built in; v1 shipped on May 21, 2025. There are desktop builds for Mac, Windows, and Linux plus a browser version at app.zoo.dev, so you can start anywhere.

The output story is clear: every Text-to-CAD call returns a STEP file, optionally with the KCL code. Generated models drop straight into the familiar sketch-and-feature-tree workflow, so “AI roughs it out, a human finishes it with dimensions” happens inside one window. You never need to write KCL: Zoo states that point-and-click covers sketching, extruding, and assemblies, with code recorded silently behind the GUI. Programmer types can edit KCL directly and mass-produce variants instead.

The architecture makes a deliberate trade. The application itself (KittyCAD modeling-app) is MIT-licensed open source on GitHub, while the geometry engine at its heart is proprietary and runs GPU-accelerated in the cloud. Your local hardware barely matters, but there is no offline mode.

Pricing starts free. The Free plan includes the full core CAD workflow, 20 minutes of Zookeeper reasoning time per month, and developer API usage worth 10 dollars per month (as of July 2026, per the official FAQ). Paid tiers lift the generation limits; amounts change with plan structure, so check the official pricing page.

Getting started is three moves: open app.zoo.dev in a browser (no install), generate one part from a prompt, edit it on the spot and export STEP. That STEP opens in Fusion or FreeCAD, so you abandon nothing. For printing, export STL from your CAD into the slicer. The browser build is officially positioned as a light trial environment, so move to the desktop app once the tool earns a place in your workflow.

Zookeeper: From One-Shot Generation to Conversational Design

The other reason Zoo defines Text-to-CAD in 2026 is Zookeeper, its conversational CAD agent. Early Text-to-CAD was fire-and-forget: one prompt, one shape. Zookeeper iterates through chat. Tell it “add two more of these holes” or “make the walls 3 mm” and the agent edits the existing model.

It is more than repeated generation. Per the official docs, Zookeeper uses engine-level tools to inspect in-progress geometry, take snapshots, and debug as it builds. It also answers physical-property questions, computing surface area, volume, mass, and center of gravity, so “how many grams of material will this design need” becomes a design-review question you can just ask. The trajectory mirrors what happened in coding: autocomplete grew into agents that complete tasks through dialogue, and CAD AI is moving from shape generation toward participation in the design process. The free 20 minutes per month is plenty to try a few flange-or-bracket-sized parts; start by ordering a replacement for something broken around your desk.

Writing Prompts: Dimension-First Language

Text-to-CAD prompts follow different rules from mesh-generation prompts. Adjectives about mood and style give way to nouns and numbers naming dimensions, quantities, and features. Zoo’s own site offers model examples: “Helium tank, 9 in x 2.5 ft, top valve, 2 clamps” and “Engine valve, 120 mm long, 30 mm head, 6 mm stem.”

The template those two examples teach: name the part in industry vocabulary, list the governing dimensions with units, then enumerate attached features with quantities. There is no room for “sleek” or “cool”; write as if compressing a drawing title block into one line. Conveniently, writing such a prompt is itself a design act, because it exposes what you have not decided yet. If you cannot state the hole pitch of the bracket you are ordering, the design conditions are unsettled regardless of any AI. Text-to-CAD doubles as a training device for verbalizing requirements.

Also worth internalizing: when the result is off, choose between re-prompting and feature editing. Wrong overall structure means rewrite and regenerate; a few wrong dimensions mean switch to feature edits. Because there is an editable exit, you can stop demanding perfection from the prompt.

Choosing Between Text-to-3D and Text-to-CAD

Two litmus tests settle borderline cases. Will you ever want to change one dimension later? If yes, the output should be B-Rep from the start. Will you hand the data to someone else? Machinists and collaborators move far faster with ISO 10303 STEP than with triangle soup. Hybrids are practical too: generate the decorative half of a phone stand with a mesh tool, build the functional slot and angle with Text-to-CAD, and merge in CAD. For picking mesh-generation tools, see the Text-to-3D SaaS comparison. A maker who handles both looks and fit should carry both tools in 2026.

Summary: Editability Is the Dividing Line

Text-to-CAD in one sentence: generative AI whose exit is engineering data instead of a mountain of triangles. Generated meshes stall on functional parts because meshes structurally cannot store dimensions as design intent. B-Rep (STEP, ISO 10303) preserves dimensions and topology, and recovery from mesh to B-Rep is impractical, so generation must start from B-Rep. Zoo Design Studio is the representative implementation: STEP plus optional KCL per generation, an MIT-licensed app over a cloud engine, and a free tier that includes 20 minutes of Zookeeper agent time monthly. The prompt discipline is part name, governing dimensions, feature quantities. And the litmus tests are “will I retype a dimension” and “will I hand this file to someone.”

Use the free tier to order a replacement for one broken part in plain language. The moment you open the generated STEP in CAD and retype a hole diameter yourself, “editable 3D model” stops being a phrase and becomes muscle memory. For how to run Zoo day to day, continue with Zoo Text-to-CAD in Practice, and for the wider tool landscape, the AI CAD copilot comparison.

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.