Just Use text-to-cad for 3D-Printable Parts

You own a 3D printer and you keep running into the same wall. You need a cable clip that fits your exact desk edge, a bracket two millimetres narrower than anything on Printables, a replacement knob for the coffee machine, a spacer with a specific M3 hole pattern. Searching model repositories burns an hour and finds something close but wrong. Learning Fusion 360 or FreeCAD will take weeks before you produce anything usable.

text-to-cad is the pick. Describe the part in plain English to an AI agent — Claude Code, Cursor, Codex, Gemini — and get back real parametric CAD: a STEP file where 12 mm means 12 mm, holes are geometrically round, and you can ask for changes in the same conversation. It is MIT-licensed, free, runs locally on your machine, and the ~20.4k-star GitHub repo ships weekly releases.

Why text-to-cad

The core problem with every other “AI 3D” tool is that they generate meshes. A mesh is a bag of triangles — it looks like a part but it isn’t one. Dimensions drift, holes are faceted polygons, wall thickness is whatever the AI sculpted. You can print a mesh, but you cannot reliably edit it or trust that a 6 mm hole will fit an M6 bolt.

text-to-cad generates boundary-representation BREP geometry through build123d and OpenCascade — the same kernel that sits under FreeCAD and many commercial CAD packages. The master output is STEP. STL, 3MF, and GLB exports come from that STEP, so every downstream format inherits the same precise geometry. When you ask for a 2 mm chamfer on the top edge, you get a 2 mm chamfer — not an approximation.

The workflow is conversation-driven. You describe the part; the agent builds it. You inspect it in the local CAD Explorer (a React/Vite browser preview that ships with the tool) and ask for changes in the same thread: “make the base 5 mm thicker”, “add two M4 through-holes 20 mm apart”. Iteration is just more prompting — no dragging handles, no sketching constraints.

Before you commit filament, the DfAM Check skill runs a printability analysis:

• Wall thickness measured by ray-cast (flags anything below process minimums)
• Overhang maps for down-facing surfaces, with angle callouts
• Support volume estimate so you know the real print cost
• Six-orientation ranking so you can pick the position that minimises supports
• Coverage across FDM/FFF, SLA/DLP, SLS, and metal PBF-LB processes

That check runs before the file ever touches your slicer. When the geometry passes, you export 3MF and hand it to OrcaSlicer, or push directly to Bambu Studio or Bambu Connect if that’s your printer. The agent does not press Print — that button stays yours.

Two other capabilities earn their place in a maker workflow. First, text-to-cad can fetch off-the-shelf STEP files for standard hardware — M3 screws, bearings, stepper motors, connectors — and embed them in the assembly so you can check fit before printing. Second, it generates dimensioned engineering drawings (PDF with views, hidden lines, hole callouts, title block) and DXF profiles for laser cutting or CNC routing. If a part starts as 3D-printed and later gets cut from aluminium plate, the DXF is already there.

On costs: text-to-cad is free. You need Python 3.11 or later, uv for package management, and one of the supported agent apps. Setup takes roughly 30 minutes the first time. The tool sends anonymous usage telemetry by default; one command disables it permanently:

uvx cadgen telemetry off

The repo updates weekly — last release was 2026-10-11 — and the MIT licence means nothing changes about your rights to use it however you want.

Who should not use text-to-cad: if you want to design organic sculpted shapes — figurines, miniatures, artistic objects with flowing surfaces — this is the wrong tool. Those need mesh generation, and Meshy exists for exactly that. text-to-cad is an engineering tool. Parts with functional geometry, mating surfaces, and specific dimensions are where it works.

Why not Zoo Design Studio

Zoo Design Studio is a serious alternative and the right pick for one specific situation: you do not want to install Python, configure an agent, or run anything locally.

Zoo runs in the browser and on a desktop app (Mac, Windows, Linux). Its Zookeeper agent is built directly into the design environment — you type a prompt, it builds the model in KCL (Zoo’s own parametric language), and you can click to edit dimensions or refine with more prompts without leaving the app. Output is true B-rep geometry: STEP, glTF, FBX, STL, OBJ. The free tier includes 20 Text-to-CAD credits per month, which is enough for occasional parts. The Plus tier is $20/month for roughly 400 minutes of Zookeeper usage.

The trade-offs are real. The free tier allows Zoo to train on your model data — there is an opt-out, but you have to find and trigger it. Plus requires a manual opt-out too. Pro ($99/month) excludes your data from training by default. For a maker printing brackets for personal use, that probably does not matter. For anyone designing parts for a product they intend to sell, read the data terms before you commit geometry to a hosted service.

The second trade-off is cost structure. text-to-cad is free forever; Zoo is free up to 20 jobs per month and then you pay. If you iterate heavily — rebuilding a part five times to get the fit right — those credits go fast.

Pick Zoo if: you want a visual CAD environment with no local setup and you can live within 20 generations per month (or budget $20/month). Pick text-to-cad if: you already pay for Claude Code, Cursor, or Codex, you want your designs to stay entirely on your machine, and you do not mind a 30-minute setup.

Why not the rest

AdamCAD generates parametric models from text prompts and the browser demo is fast and accessible. The problem is output format: AdamCAD exports STL, OBJ, and OpenSCAD (.scad) files. STEP export is listed as “coming soon”. OpenSCAD is a scripting approach to CAD, not native B-rep — it works, but the geometry is less interoperable with downstream tools. If you need a STEP file to send to a machine shop or import into an assembly, AdamCAD cannot deliver that today.

Meshy does one thing well: it turns a text prompt or a photo into a textured 3D mesh in under two minutes. Game assets, character models, concept art for a product pitch — Meshy is fast and the results look good. For 3D printing a functional part, it is wrong by design. A 10 mm hole in a Meshy model is whatever shape the mesh generator produced, not a 10 mm cylinder. Meshy has a printability check that flags obvious mesh problems, but it cannot give you dimensional accuracy that was never there. Start in Meshy if you want a figurine; start in text-to-cad if you want a bracket.

Tinkercad is how most people learn 3D modeling, and Autodesk has made it genuinely good for that purpose. The problem is that it is manual modeling — you drag shapes, set dimensions in a sidebar, combine and subtract primitives by hand. That is exactly the workflow you are trying to avoid. Tinkercad does not take a text description and build geometry; it gives you a browser-based CSG editor. It is free and approachable, but it is not the tool for this problem.

When to revisit

text-to-cad is built around AI agent conversations, not a visual editor. If you find yourself wanting to click and drag — adjust a fillet by eye, position a hole by feel — the tool will frustrate you. At that point, Zoo Design Studio’s hybrid UI (prompt plus point-and-click editing) is worth trying before committing to a full parametric CAD application.

If you grow into designing parts for sale and need formal version control, release management, and team collaboration on geometry, look at what text-to-cad’s STEP output can feed into — Onshape handles multi-user parametric CAD without a desktop install and accepts STEP import. That is a different problem from printing parts at your desk today, but it is where the workflow naturally leads.