Text-CAD: It’s not about “generating a part drawing.” It’s the full pipeline: one sentence → STEP → slicing → directly connecting to a local Bambu printer for output.

Previously, mechanical engineers manually modeled in SolidWorks and then checked each dimension with calipers—this project takes over both actions.

It bundles 11 skills at once, covering the entire hardware pipeline from design to manufacturing:

1. CAD: Natural language / images / 2D drawings → STEP, with export to STL/3MF/GLB
2. step.parts: Automatically find ready-made STEP files for screws, bearings, motors, and connectors
3. DXF: Generate 2D cutting layouts; SendCutSend: Validate before uploading for laser cutting
4. G-code + Bambu Labs: Slice → dry-run on local printer → print cautiously
5. URDF/SRDF/SDF: Write robot structures, MoveIt2 planning, and simulation worlds

The key design choice is clever: instead of outputting STEP binary, the model writes build123d Python source code, then runs gen_step() to generate it. This gives the LLM a text-based intermediary it’s truly good at, bypassing the deadlock of “directly generating a hard format”—same logic as the [[video-use]] approach we’ve discussed: video-use makes the LLM read structured text instead of video, text-to-cad makes the LLM write Python instead of touching the geometry kernel. Both give the model a high-value token landing point.

Seen side by side with Jellyfish, both are about harnessing engineering—one in AI short videos, the other in hardware manufacturing:

• Jellyfish treats “character consistency” as a first-class problem, turning uncontrollable generation into traceable processes.
• Text-to-cad treats “geometric correctness” as a first-class problem—scripts/inspect performs refs/facts/planes/measure/align/diff validation, visual verification with snapshots is mandatory (deterministic checks passed or not, you can’t skip), and on failure: “edit the minimal responsible source segment → regenerate → rerun failed items.”

What’s the cost?

It makes each skill a “thin wrapper + progressive references”—only loading the reference documents needed for the current task, not the whole set. This saves context, but it also means the ceiling depends on the underlying CLI (build123d / slicer / OCCT kernel). The model only orchestrates, it doesn’t handle geometric solving.

#TextToCAD #AICAD #AgentSkills #build123d #HardwareAI

step.parts | 16,000+ open source STEP parts for your next CAD project

Source: https://www.step.parts/
Open Source CAD Directory

Loading catalog