How Text to CAD works
Text to CAD is our prompt-to-STL workflow for precise, editable 3D models. It uses generated CadQuery Python as the bridge between natural language and real CAD geometry.
Short answer
MakeIt3D Text to CAD turns a written prompt into CadQuery code, runs that code in a CAD worker, previews the generated model, and exports an STL. The core advantage is editability: the model is not only a mesh, it starts as readable parametric source code.
Research basis
CadQuery is a Python library for building parametric CAD models. Its documentation emphasizes plain-text, editable source code, customizable parameters, and CAD exports such as STEP, AMF, 3MF, and STL. It is built on OCP, the Python binding layer for the OpenCascade modeling kernel.
That makes CadQuery a practical target for AI generation: the output is code, not a black-box mesh. The generated script can expose dimensions as variables, express design intent, and be rerun after edits. CadQuery's own design principles also align with this: scripts should read close to how a person describes a part and should capture design intent.
OpenAI's agent tooling points to the next layer of improvement. Agents can use tools, execute code, call functions, retrieve files or docs, and hand work to specialized agents. For CAD, that means the generator does not have to be a single pass. It can become a loop: plan, write, execute, inspect, repair, validate, and explain.
Current pipeline
The current version is intentionally simple: generate code, run it, preview the STL, then let the user iterate.
Prompt to design intent
The system first extracts dimensions, constraints, part families, output format, and likely manufacturing intent from the user's prompt.
Design intent to CadQuery
The model writes Python that builds a parametric shape with CadQuery workplanes, sketches, selectors, booleans, fillets, shells, and exports.
Worker execution
A CAD worker runs the script in a constrained environment, using CadQuery and the underlying OCP/OpenCascade modeling kernel to create geometry.
Preview and export
The app receives the generated STL, previews it in the browser, and lets the user download the model for slicing, prototyping, or iteration.
Why this matters
Text prompts are good for intent, but CAD requires exact geometry. CadQuery gives the system a precise code target.
Source-code CAD is debuggable. If a fillet fails, a hole lands on the wrong face, or an export breaks, the script can be repaired.
Parameters can remain visible. A user can change width, wall thickness, clearance, or hole count without regenerating from scratch.
The same workflow can later export richer CAD formats, not only STL, because the underlying model starts as CAD operations.
Agent skills roadmap
In this context, agent skills mean narrow, reusable abilities around the CAD workflow. They would let Text to CAD behave less like one prompt and more like an engineering assistant.
Part-family skill
Classify prompts into brackets, boxes, gears, adapters, furniture, architectural blocks, or decorative objects, then apply specialized templates and constraints.
Dimensional reasoning skill
Normalize units, infer missing dimensions, flag impossible dimensions, and keep key parameters editable at the top of the generated script.
CadQuery repair skill
Read stack traces and geometry failures, patch the script, rerun it, and stop only after the STL export succeeds or the request is clearly impossible.
Manufacturability skill
Review wall thickness, overhangs, hole clearances, tolerances, watertightness, and slicer-friendly orientation before presenting the final model.
Documentation retrieval skill
Pull the relevant CadQuery examples and API references for the requested operation, reducing hallucinated method names and improving code quality.
Evaluation skill
Render thumbnails, inspect bounding boxes and mesh statistics, compare the output against the prompt, and produce a short quality report.
Theoretical upgrade
Multi-agent CAD loop
A planner agent could translate the prompt into a part specification. A code agent could write CadQuery. A validator agent could run the script, inspect mesh metadata, and compare the result against the prompt. A repair agent could patch failures until the model exports.
Safer execution boundary
The CAD worker should stay sandboxed. Agent skills should not mean arbitrary execution. They should mean better allowed operations, clearer validation, smaller tool scopes, and repeatable checks before returning a downloadable file.