A German toolmaker-turned-software-developer is asking makers to try, and break, a browser tool that promises to turn a plain-language description into a printable part. In a post on the Prusa3D forum dated September 22, Michael Bender introduced Constructor AI and said he was looking for real-world testers. Trade coverage followed within days: TCT Magazine reported the free public test on September 25, and 3D ADEPT published its own write-up on September 29.

The pitch is familiar to anyone who has watched the recent wave of text-to-3D demos. The details are not. Constructor AI is not aimed at sculptural, decorative output. Bender, who describes himself as a trained toolmaker and computer scientist, is targeting functional parts: brackets, adapters, replacement parts, and features such as holes, countersinks, threads and fits. That is the territory where most hobbyists reach for CAD, and where most of them would rather not.

How the workflow is meant to run

According to Bender's forum post, the user describes a part in normal language. The tool then asks for whatever dimensions are missing, generates the geometry, and displays the result in a 3D viewer. The interaction is a question-and-answer loop rather than a single prompt that returns a finished model, which is a sensible design choice for parts that have to fit something real. A bracket with an unspecified hole spacing is not a bracket; it is a guess.

TCT's report describes the same flow as guided step by step. Users describe the part and its dimensions, and the software walks them through the rest. The tool supports German and English, per the forum post, and a free demo is hosted at constructorai.eu.

Beyond generating new parts, the tool can import and modify existing files. TCT says it handles both ASCII and binary STL, and that models can be extended, modified and exported again as STL files. That matters for a practical reason: a great deal of functional printing is modification work, such as adding a mounting hole to a downloaded part or bolting an adapter onto something that already exists.

From geometry to G-code

The feature that separates this from a pure modeling toy is the output end of the pipeline. TCT reports that the tool takes printing parameters and material into account and generates G-code from the finished model, with extra functions available for supported printers. 3D ADEPT says the project's website lists Kobra 2 and Kobra S1 workflows, though it notes those compatibility claims have not been independently tested; neither report explains which slicing settings are exposed or how the G-code is produced. Those are open questions for anyone planning to send output straight to a machine. Makers should treat any G-code from an early test build the way they would treat G-code from any unfamiliar source: preview it before printing.

The tool also handles multiple workpieces in one session, with positioning and rotation, along with holes, connections and guides, according to TCT. Image, logo and lithophane functions are described as still being tested, which places them firmly in experimental territory rather than the core feature set.

What an early tester found

The most useful data point so far comes not from the announcement but from the replies. A forum user named Jürgen tried the free demo and reported that the AI follows a strict, fixed course and fixed terminology. There was no way to ask the system to clarify what it meant. His example was the word "thickness," which can be ambiguous when the part is a cylinder and the question could refer to its height. He also reported that a formula entered as a wall-thickness value was not accepted, and suggested that a language model working in the background would let users ask clarification questions.

He described further trouble: after unexpected answers, later inputs stopped being accepted, and he could not find a way to specify the position of a cross hole. He called both of those looking like bugs, and the rigid dialog a missing feature.

That critique cuts to the central tension in this category. A guided, deterministic flow is predictable, but it puts the burden of learning the tool's vocabulary on the user. That undercuts the "plain English" promise. A free-form language model is friendlier to talk to, but it can be vague about exactly the numbers a functional part depends on. Jürgen's suggestion points toward a hybrid: keep the structured dimension prompts, but let a language layer translate a maker's phrasing into the tool's terms.

Whether Bender adopts that approach is not something the sources say, and this is precisely the kind of feedback the test appears designed to collect. TCT notes that testers can give feedback and report problems through integrated live support.

A test, not a launch

Bender has been direct about the maturity of the project. TCT reports that he stresses the test is not intended to give the impression of a finished product, and 3D ADEPT says the same. That framing is worth taking at face value. A public test with live feedback support is a way to find where real users get stuck, and the forum thread suggests it is already doing that.

It also means makers should calibrate expectations. The available reporting describes features and intentions, not benchmarks. There is no published accuracy data, no comparison against conventional CAD workflows, and no account of how often the guided flow produces a part that prints correctly on the first try. Those numbers do not exist in the sources we reviewed, and we have not run our own tests, so we are not offering any.

What It Means for Makers

For makers who design occasionally rather than daily, the appeal is obvious. Many functional prints are simple: a spacer, a clip, a bracket with two holes. Learning parametric CAD for that class of part can feel disproportionate. A tool that asks for the dimensions and hands back an STL and G-code addresses that gap directly. The STL import and edit function may prove as useful as the from-scratch generation, since modifying an existing part is often faster than modeling one.

For experienced CAD users, the value proposition is thinner today. Rigid terminology, no clarification mechanism and a rejected formula input are the sort of limits that a parametric modeler with variables and expressions does not have. That comparison is our framing, not a finding from the sources, but it follows directly from what the early tester reported.

The practical advice is simple. The test is free and runs in the browser, so the cost of trying it is a few minutes. Start with a part whose dimensions you already know, verify the geometry in the viewer, measure the exported STL in your slicer, and inspect the G-code before committing filament. If you find friction, the developer says he wants to hear about it.

The larger question is whether an independent developer can make a guided flow feel natural enough to replace CAD for everyday parts. Constructor AI's early public test suggests the idea has attention, and the first forum feedback suggests the execution has work left to do. Both can be true of an early test.

Sources

Correction, Oct. 22, 2026: An earlier version of this article said 3D ADEPT noted the release does not say which printers are supported. 3D ADEPT says the project's website lists Kobra 2 and Kobra S1 workflows, and that these compatibility claims have not been independently tested.