Can AI generate editable CAD models?

First decide what editable means

An image generator can show an object without producing any usable geometry. A mesh describes a surface; it is a common deliverable for 3D printing but does not preserve the original design parameters. A CAD interchange file serves another purpose: carrying model geometry between tools. Do not accept an attractive image as a substitute for the file your workflow needs.

CadQuery documents the important limit directly: its supported interchange formats do not carry its parametric data. The fully parametric representation is its own Python source [2]. Exporting STEP is useful for CAD interchange, but it does not promise that another application will recover the original feature history or parameter relationships. Keep both source and export.

• Need repeatable dimension changes? Ask for named parameters and the native script.
• Need geometry in another CAD application? Confirm its import support before choosing the export format.
• Need a 3D-printing handoff? Request a supported mesh and inspect it in the actual slicer.
• Need only a concept illustration? Say so; that is a different task from CAD.

How OpenSCAD and CadQuery fit the workflow

OpenSCAD reads a script that describes an object and compiles it into a 3D model. Its official overview describes constructive solid geometry and extrusion of 2D outlines [3]. This gives a coding assistant a text representation it can revise, while the parameters remain available for a person to inspect.

CadQuery uses Python to construct parametric models and supports operations on an OpenCascade-based geometry representation [1]. Its documented STEP export is useful when the deliverable must move into another CAD or analysis system [2]. Choosing it also means reviewing generated Python as executable code, not treating it as a harmless drawing file.

RepoRadar recommendation: choose by the artifact and checks you require, not by which syntax a model claims to prefer. A source-controlled OpenSCAD design can suit simple parameterized solids; a CadQuery workflow can suit Python-based geometry checks and CAD interchange. Neither removes the need for design review.

What the current AI experiment actually establishes

ModelRift reports testing one agent for each combination of three tasks and two toolchains: a bracket, a two-part enclosure and a threaded adapter. The same model drove all six runs. Its report says both toolchains ultimately produced clean exported meshes, after iterations and independent file checks [4]. Those are the experimenter's findings, not RepoRadar hands-on results.

The useful warning is that plausible renders and successful tool status messages missed defective geometry. The report describes an apparently correct threaded part with a missing internal core, detected by a volume measurement. It also reports an OpenSCAD export whose mesh defects contradicted the tool's clean status [4].

Do not turn this small experiment into a universal leaderboard. The authors disclose one run per cell, differences in preview tooling, exclusion of the BOSL2 library, and a focus on construction rather than repairing someone else's model. ModelRift also sells an OpenSCAD-based service. The evidence supports adding verification, not declaring that either tool always wins.

Write a specification the output can be checked against

Start with a simple non-critical object you can measure. State units, overall dimensions, hole positions, wall thickness, required clearances and the intended output format. Identify unknowns instead of asking the model to guess how a mating part is shaped.

Separate design choices from acceptance checks. For example, a prototype box may need a named cavity width and a separate clearance parameter. Ask for those values in the source and then check the exported cavity, rather than assuming that a parameter name proves the geometry uses it correctly. This is a suggested workflow, not a tested design specification.

• Request the native script, exported file and a short list of unresolved assumptions.
• Change one dimension and regenerate to check whether dependent dimensions update coherently.
• Keep versions so you can compare changes and return to an earlier design.
• Require the agent to report failed checks rather than silently changing the specification.

Check the export, then check the real use case

Use an independent viewer or geometry tool to inspect the exported file. Check scale and bounding dimensions, openings, disconnected parts and unintended intersections. For a mesh, examine watertightness and non-manifold edges; a clean mesh still does not prove that clearances, wall thickness or loads are correct. ModelRift's examples make that separation concrete [4].

Inspect cross-sections when exterior renders hide internal geometry. Preview the file in the intended slicer before making a low-risk test print. If parts must fit together, measure the fabricated result; material, manufacturing settings and tolerances remain outside what a screenshot can establish.

Do not use an AI-generated model as sole engineering sign-off for load-bearing, pressure-containing, medical, electrical-safety or other critical parts. Qualified design review and appropriate physical testing are separate requirements.

Keep the automation boundary narrow

Run generated scripts in a disposable project with limited file access. Do not give a CAD agent unrelated credentials, private documents or unrestricted command execution. If you send proprietary drawings to a hosted model, first verify that the account and data policy permit it.

Next action: take one simple non-critical part, write its required dimensions and deliverables, and ask for a parameterized source file. Change a parameter, regenerate, then inspect the export independently. Success is an editable artifact that passes your checks, not an assistant saying the job is finished.