Benchmark: CadQuery vs. OpenSCAD for agentic CAD work

Source: https://modelrift.com/blog/cadquery-vs-openscad

Author: ModelRift

Published: 2026-09-12

Overview

ModelRift generates OpenSCAD for every model on its platform, and periodically re-tests that choice. This controlled comparison asked a narrow question: which tool can an AI agent drive to a correct, printable, functional part *with nobody watching*? Six agents (one per cell, no cross-talk), three tasks, two tools, and every resulting STL checked by an independent parser that trusts neither tool.

Setup

All six runs driven by Claude Opus 5 (1M context) through Claude Code. Tool versions: CadQuery 2.8.0 on Python 3.14 vs OpenSCAD 2026.06.12. The OpenSCAD side ran on ModelRift's published openscad-skill; for CadQuery they ported that skill operation-by-operation (adding an offscreen renderer and B-rep validity metrics), with identical 3D-printing design rules on both sides. Tasks: T1 L-shaped shelf bracket; T2 two-part snap-fit enclosure for a 50x26 PCB that must actually fit; T3 M24x2 threaded hose-barb adapter with a true helical thread — stacked rings explicitly banned.

Results

All six parts came out printable and clean. Capability was the boring part of the answer. Iteration count was identical (11 versions each). What differs is failure character:

CadQuery fails loudly and early. Error messages are poor (BRep_API: command not done names neither edge nor radius), but an exception stops the run — and a stopped model cannot ship by accident. In T1 the agent had to bisect by hand to discover two R3 fillets don't fit a 4 mm wall.
OpenSCAD fails silently and late. In T2 it reported zero errors/warnings across ~45 invocations while deleting mounting posts, misplacing slots, and certifying a corrupted export as clean; its Manifold backend stamped Status: NoError on an STL with 4 non-manifold edges and 60 zero-area triangles.
T3 upset: OpenSCAD finished in one version, correct on first compile, in 43 ms — hand-writing the helix as raw vertex/face arithmetic in a single polyhedron. CadQuery's readable makeHelix+sweep worked, but union() silently discarded the core cylinder because the thread root sat exactly on the core radius (volume 7065 mm³ vs 10323 expected; valid=True, solids=1). Only a volume measurement caught it.
Renders caught nothing that mattered. Across six runs, images caught only coarse blunders; every print-ruining defect was found by a number — volume, angle, interference test.
CadQuery can be interrogated; OpenSCAD cannot. Agents asserted box/lid interference volume and read countersink angles off the B-rep; OpenSCAD agents independently wrote their own STL parsers to measure what they had built.

Takeaways

Verifiability beats expressiveness in an agent loop: being able to ask the kernel what you just built — and assert on the answer — outweighs syntax convenience. Task shape decides more than the tool. If a pipeline leans on screenshots, it leans on the one channel that caught nothing here; force numeric assertions (wall thickness, clearance, interference volume, overhang angle) and audit the exported mesh independently, because both toolchains certified wrong geometry. ModelRift stays on OpenSCAD, adding a feature-edge renderer and a mandatory numeric check pass.

中文概要

ModelRift 用同一模型（Claude Opus 5 1M，经 Claude Code）驱动六个独立子智能体，在三个 CAD 任务（L 形支架、50x26 PCB 卡扣外壳、M24x2 螺纹软管接头）上对比 OpenSCAD 与 CadQuery，所有 STL 由独立解析器审计。六个零件全部可打印，迭代次数相同（各 11 版），真正的差异在失败模式：CadQuery "响亮而早"——报错难懂但会停机；OpenSCAD "沉默而晚"——T2 任务约 45 次调用零报错，却删掉了安装柱、导出损坏 STL 还自证干净（Manifold 后端把含 4 条非流形边的网格判为 NoError）。最大冷门在 T3：OpenSCAD 手写 polyhedron 真螺旋螺纹一次编译通过（43ms、1 版），CadQuery 的 union() 因螺纹根与核心柱恰好相切而静默丢掉核心柱（体积 7065 vs 10323 mm³，valid 仍为 True），靠体积测量才被抓到。渲染图六次运行没抓到任何致命缺陷，救场的全是数字。结论：对无人值守生成，可验证性压过表达力——必须强制数值断言（壁厚、间隙、干涉体积、悬垂角）并独立审计导出网格。