ModelRift recently evaluated two code-first CAD tools, CadQuery and OpenSCAD, to assess their effectiveness when driven autonomously by AI agents in generating functional 3D-printable parts. The study focused on whether an AI could produce correct models without human intervention, rather than on ease of manual coding.
The experiment involved six AI agents, three design tasks, and both CAD tools. Each AI agent worked independently, using Claude Opus 5 through Claude Code, with no communication between agents. The CAD versions tested were CadQuery 2.8.0 on Python 3.14 and OpenSCAD 2026.06.12, running on an M-series Mac.
The tasks ranged from a simple L-bracket shelf mount to a two-part snap-fit enclosure for a PCB, and a complex M24x2 threaded hose-barb adapter requiring true helical geometry. Both tools produced printable STL files for all tasks, but their failure modes and verification approaches differed significantly.
OpenSCAD completed the helical thread task quickly by manually coding the helix as a polyhedron, despite lacking native sweep functionality. CadQuery used its B-rep kernel to model the thread more succinctly but encountered silent failures during boolean operations that compromised the geometry without immediate error messages.
Verification proved critical. Both tools’ internal validity reports were insufficient, prompting ModelRift to implement independent mesh parsing to check for watertightness, manifold edges, and volume accuracy. This revealed that OpenSCAD sometimes certified flawed meshes as error-free, while CadQuery’s exceptions prevented silent failures but provided less informative error messages.
CadQuery’s ability to query and assert geometric properties during model generation gave it an advantage in unattended workflows, enabling early detection of design violations. OpenSCAD relied on textual echoes, which require human interpretation, making automated verification more challenging.
Performance-wise, OpenSCAD was significantly faster in geometry recomputation, but this was less relevant in AI-driven loops dominated by inference time. Visual rendering differences also emerged; OpenSCAD’s renderings lacked clear part edges, while CadQuery’s B-rep allowed for precise edge detection, influencing agent feedback quality.
ModelRift concluded that neither tool outperformed the other in output quality, but verification capabilities and failure transparency are decisive factors for autonomous design. The study highlighted the importance of numeric assertions and independent mesh validation over visual inspection alone in AI-driven CAD workflows.
Despite CadQuery’s verification strengths, ModelRift will continue using OpenSCAD due to its compact syntax, sandboxed environment, and rapid rendering, while planning to enhance OpenSCAD’s verification with additional numeric checks and improved rendering for agent feedback.
This comparison underscores the evolving requirements for CAD tools in AI-assisted design, emphasizing robust verification and clear failure reporting to support reliable autonomous model generation.