Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableHabr AI→ original

Text-to-cad: AI Agent Draws Robot Parts From a Description, Project Passes 26,000 Stars

CAD, the software used for 3D drawings, is hard to pick up without practice. Mechatronics engineer Jake Fitzgerald handed this work to Codex and Claude Code: the agent writes a script, builds the model, looks at a snapshot, and fixes errors. The open-source text-to-cad project has about 26,000 stars (developer bookmarks) and half a million downloads. The takeaway: the agent speeds up simple parts, but exports and dimensions must be checked separately.

AI-processed from Habr AI; edited by Hamidun News
Text-to-cad: AI Agent Draws Robot Parts From a Description, Project Passes 26,000 Stars
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Scripts in Gitunder 3 MB of code produces about 10 GB of models
Key metric #2A faulty DXF export from the workshop delivered parts with elongated holes
Key metric #3The agent found the cause of arm wobble and added a ring, which lowered stresses

To design a robot part, you need CAD, the software for 3D drawings, and it is not easy to master. Mechatronics engineer Jake Fitzgerald, who wrote software for many years and forgot CAD, gave this work to the Codex and Claude Code agents. They write code, the code builds a 3D model, and then the agent looks at a snapshot and fixes errors. This grew into the open-source text-to-cad project: about 26,000 stars, meaning interest marks from developers, and half a million downloads. At Ubuntu Summit 26.10, Fitzgerald described how he uses it to build an inexpensive six-axis manipulator and where things break down.

How text-to-cad works

Fitzgerald is a mechatronics engineer by training. He then worked in software development, including at WhatsApp, and became a co-founder of a fintech company. Returning to robotics, he set the goals: six degrees of freedom, a 600 mm reach, a 500 g payload, components costing less than $200, and assembly in 20 minutes. He decided to make the housing from sheet metal for rigidity. But after ten years without CAD, working in Onshape proved difficult.

So he gave the task to Codex. A local MCP server and a set of skills, Markdown instructions describing which commands to use and in what order, are installed in Claude Code or Codex. The agent writes a script in build123d, a library built on the open Open CASCADE kernel. Then it builds the model, takes a snapshot, assesses whether the shape matches the task, and repeats the cycle. For dimensions, separate geometric checks can be added.

Feedback is convenient too. A face, edge, or part is selected in the viewer, and a link with a unique identifier is inserted into the prompt. You can write "move this hole to the right and double it," and the agent understands exactly which one. It also finds the line of code that built the feature, which saves time when a script grows to tens of thousands of lines. If something is hard to explain in words, you can draw annotations over the snapshot. Keeping the geometry in memory sped up rebuilds: previously even a minor edit could cost two minutes waiting for a STEP export.

Where agents go wrong

The author does not hide the problems. Models struggle to hold onto requirements: you ask for a bracket from sheet metal and get a multilayer, distorted geometry that cannot be manufactured that way. Errors accumulate as the assembly grows more complex. No model has yet designed a complete manipulator in a single run.

The most costly errors surface in production. The agent exported the DXF outline incorrectly, and after a week of waiting, parts arrived from SendCutSend with elongated holes. On screen, everything looked plausible.

The contrast is also telling. The Fable 5.1 model produced a concept of a complex car engine in a single run over about two days, and there are many such examples in the training data. An unusual sheet-metal bracket has no such precedent, and even Opus 5.5 gave only a small improvement. Progress is still noticeable: with the Astra model, according to the author, spatial reasoning improved considerably. The agent performed a modal analysis of a swinging arm, suggested removing part of the material and adding a stiffening ring, after which the calculated stresses dropped.

What else can be automated

Beyond geometry, agents take on adjacent tasks:

  • calculations: sending the model to the open CalculiX solver and comparing variants with the finite element method;
  • drawings: dimensions, tolerances, and hole tables, which sometimes take engineers as long as the model itself;
  • manufacturability checks: a skill reads SendCutSend's requirements and warns when a hole is too close to a bend or when a sheet is unsuitable for the process;
  • 3D printing: Jake described a camera bracket from his phone at the gym, and by the time he got back the part was already on the printer bed;
  • control: Claude wrote the software side in about half an hour.

The main practical advantage is that CAD becomes code. Scripts live in Git, changes have a history, and in the examples directory under 3 MB of code produces about 10 GB of models. When parameters change, a URDF is rebuilt: a file describing links, joints, and their motion limits. It is used to check inverse kinematics and the reachability of points in the workspace. For example, you can reduce the shoulder range by 20° on each side and find out how many points become unreachable.

Forecasts and limitations

The author expects engineers to run dozens or hundreds of agents with slightly different requirements and pick the best variants. For now, thousands of parallel runs over several days are too expensive. He also believes a language model will be able to describe an increasing range of outputs in code, up to controlling a robot's end effector. These are still hypotheses, not results.

For practitioners, the conclusion is this: the agent is already useful as an accelerator for simple parts, calculations, and drawing preparation, but geometric checks and export control must not be removed from the pipeline. The project remains experimental, and a concept image still needs to be confirmed with engineering calculations and testing.

Story

Key Questions Answered

How text-to-cad works?

Fitzgerald is a mechatronics engineer by training. He then worked in software development, including at WhatsApp, and became a co-founder of a fintech company. Returning to robotics, he set the goals: six degrees of freedom, a 600 mm reach, a 500 g payload, components costing less than $200, and assembly in 20 minutes. He decided to make the housing from sheet metal for rigidity. But after ten years without CAD, working in Onshape proved difficult. So he gave the task to Codex.

Where agents go wrong?

The author does not hide the problems. Models struggle to hold onto requirements: you ask for a bracket from sheet metal and get a multilayer, distorted geometry that cannot be manufactured that way. Errors accumulate as the assembly grows more complex. No model has yet designed a complete manipulator in a single run. The most costly errors surface in production. The agent exported the DXF outline incorrectly, and after a week of waiting, parts arrived from SendCutSend with elongated holes.

What else can be automated?

Beyond geometry, agents take on adjacent tasks: calculations: sending the model to the open CalculiX solver and comparing variants with the finite element method; drawings: dimensions, tolerances, and hole tables, which sometimes take engineers as long as the model itself; manufacturability checks: a skill reads SendCutSend's requirements and warns when a hole is too close to a bend or when a sheet is unsuitable for the process; 3D printing: Jake described a camera bracket from his phone at the…

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…