Start with the fact, because it deserves to be taken seriously.
On April 22, 2026, an open-source project called text-to-cad went live under the MIT license, written by Jake Fitzgerald. The next day its demo spread: using GPT-5.5 with Codex, he never opened a CAD program — he talked a 7-DOF robot arm into existence in plain language. A complete URDF with working kinematics, a custom GUI with live joint sliders, a full set of STEP parts and the assembly. Everything but the gripper, generated by AI. His own note: the same job used to take him weeks, stitching half a dozen tools together.
A month later a company called MecAgent posted another demo: GPT-6 Astra driving a SolidWorks 2026 harness, one prompt, a 35-minute run, a robot arm — 11 parts, one main assembly, parametric with a feature tree, editable.
Both deserve attention, because neither is “AI made a nice 3D picture.” What they hand over is STEP — the industry-standard CAD exchange format, carrying real geometry, dimensions, and assembly relationships. The deeper shift is the idea itself: the AI no longer draws the model; it writes parametric Python (build123d), and the OpenCascade kernel computes the solid. The model becomes code — readable, editable, version-controlled. That step is real progress.
1. What it hands over is a model, not a part
From “AI can draw it” to “a part that can be made” there are two more layers, and they are not the same layer.
Layer one: it looks right. The model spins in the browser, the joints simulate, the assembly rotates — this layer text-to-cad already does well.
Layer two: it goes together. This is where its own benchmark parts fell first. Someone testing the project’s built-in L-bracket benchmark found that the triangular reinforcement rib overlapped the bolt holes — as generated, the part cannot be assembled at all. That is not a rendering problem. It is an assembly problem: the rib collides with the bolt head, the washer, even the wrench clearance needed to turn the screw.
Layer three: it can be made. Harder still. MecAgent says it plainly: the model is parametric, has a feature tree, is editable, but is not fully constrained — and is “still far from something truly manufacturable.” The text-to-cad documentation does not hide it either: no tolerances, no FEA, no DFM analysis; complex freeform surfaces are a weak spot; large multi-part assemblies are unproven.
So the conclusion is blunt: AI has solidly cleared layer one, its own benchmarks have not cleared layer two, and it admits layer three is not there yet.
2. Going together is harder than looking right
Why “goes together” is so hard deserves its own paragraph. Because “looks right” on a screen and “goes together” in the physical world are two different things.
- Interference and clearance. When AI generates geometry it thinks about that one part; it does not inherently know how much room the neighboring part, the bolt, or the wrench needs. A rib crushing a bolt hole is the textbook case.
- The tolerance stack. Every dimension on a drawing carries a tolerance. Five links of ±0.1 in a chain add up to a half-millimeter of clearance gone. AI outputs nominal geometry — no tolerances, and therefore no such thing as a stack. The half millimeter that looked generous on the drawing is eaten the moment tolerances stack.
- Datums and fits. A bearing seat needs H7/h6; a locating face needs a datum. These are not questions of how pretty the model is. They are questions of whether the parts go together once made.
Those three are exactly what a machine shop handles every day. A flawless model reaches the shop floor and the first thing that happens is it gets taken apart: which faces need precision and which can be whatever, then the tolerance stack and the process sequence get rebuilt.
3. Once geometry gets cheap, the scarce thing is judgment
Zoom out on the timeline and something is quietly changing.
Drawing used to be slow. An engineer had to walk the shape through his head first — which cutter for this edge, how to drill that hole, how to hold this face. He was slow, but the act of drawing was itself a process review: the hand moved, and so did the mind.
Now the cost of generating geometry is heading toward zero. One sentence, a few minutes, a pile of candidate models. So the bottleneck moves from “cannot draw it” to judging it — deciding which model in the pile can actually go together, get cut, and pass inspection.
Whoever sits at that bottleneck is worth more. That is precisely the machine shop’s position: we may not invent shapes better than the AI, but every day we pay a real price for the question of whether a given shape can land on this machine, in this material, at this price — and we turn that price into experience.
4. So this is how we use it
No hedging.
AI diverges — it spreads one idea into dozens of possible geometries in minutes. We converge — we narrow that pile to the one part that goes together, cuts, and passes inspection.
The two ends do not compete; they complete each other. Drop either one and the part never gets made. So the stronger the AI gets, the more valuable the person who can spot, at a glance, that this one will not assemble and that one will not cut.
If you are holding a part an AI drew for you and want to know what it really costs to make it — send it over. We will not just answer “yes, we can.” We will tell you which feature is expensive, on which operation, and where changing it saves money.
Key References
- Open-source project text-to-cad (GitHub: earthtojake/text-to-cad; MIT; released 2026-04-22; author Jake Fitzgerald): the AI writes parametric build123d Python, the OpenCascade kernel computes the solid, and STEP is the primary output (plus STL/DXF/URDF and more). Built-in benchmark parts include a calibration block, a round flange, an L-bracket, a centrifugal impeller, and a planetary gear set.
- Jake Fitzgerald’s public demo, 2026-04-23 (X/Twitter @earthtojake): GPT-5.5 with Codex generated a 7-DOF robot arm’s URDF, kinematics, custom GUI, and STEP assembly — everything but the gripper generated by AI.
- MecAgent public demo (LinkedIn): GPT-6 Astra with a SolidWorks 2026 harness, a single prompt, a 35-minute run generating a robot arm (11 parts, one main assembly); the post notes the model is not fully constrained and is “still far from something truly manufacturable.”
- User testing (Hacker News discussion): a test of the project’s L-bracket benchmark found the reinforcement rib overlapping the bolt holes so the part could not be assembled; a 7-DOF arm benchmark’s joint limits were questioned as a hallucination.
Where we stand
We have made precision structural parts for twenty-three years, with thirty CNC machines including true 5-axis. The parts on a robot arm — joint housings, flanges, reducer planet carriers, brackets — are exactly what we make. They look unremarkable; misalign them slightly and they fail.
Where we can help is concrete: turning an AI-generated model into a drawing that can actually be machined (datums, critical tolerances, manufacturability), working out the tolerance stack and assembly clearance behind nominal geometry, and then cutting the small-batch prototype. If the part is stuck on “will not assemble” or “will not cut,” it is probably not the idea — it is the machining. Send us the file and we will get it right together.