When AI Agents Learn to Draw Part Drawings: text-to-cad Takes LLMs from Digital to Physical
> Project: earthtojake/text-to-cad > Language: Python | License: MIT > Today's stars: +75
---
1. The "Physical Blind Spot" of AI Agents
Over the past two years, the AI Agent ecosystem has exploded — writing code, editing docs, searching the web, managing calendars. But nearly all agents operate within an implicit boundary: they only manipulate digital objects. Code, text, images, spreadsheets — all bits, no atoms.
Ask an AI Agent to "design a bracket with three M4 screw holes and 2mm wall thickness," and it stalls. Not because it can't understand mechanical design, but because it has no hands — it can't turn a design into a STEP file for a CNC shop, or generate G-code for a 3D printer.
earthtojake/text-to-cad fills exactly this gap. It's not a text-to-image model, nor a diffusion model generating 3D shapes — it's a set of Agent Skills enabling any AI coding assistant (Claude Code, Cursor, Copilot) to generate, inspect, source, slice, and hand off CAD and robot description files.
---
2. 12 Skills Covering the Full Manufacturing Pipeline
| Skill | Input | Output | What it does |
|---|---|---|---|
| text-to-step | Text description | STEP file | Generates ISO 10303 standard CAD models |
| text-to-stl | Text description | STL file | Generates 3D printing meshes |
| text-to-dxf | Text description | DXF file | Generates 2D laser-cutting/waterjet drawings |
| text-to-urdf | Text description | URDF file | Generates robot descriptions (ROS standard) |
| text-to-sdf | Text description | SDF file | Generates simulation descriptions (Gazebo standard) |
| text-to-gcode | Text description | G-code | Generates CNC/printer instructions |
| inspect-step | STEP file | Inspection report | Verifies geometry is manufacturable |
| inspect-stl | STL file | Inspection report | Checks mesh quality (manifold, normals) |
| source-parts | BOM list | Purchase links | Quotes from SendCutSend and other suppliers |
| slice-stl | STL file | G-code | Slicing (Bambu Lab / Cura) |
| handoff-cad | CAD file | Handoff package | Packages files for manufacturers or next stage |
| dfam-check | CAD file | DfAM report | Design-for-Additive-Manufacturing check |
The last two are the most notable.
handoff-cad solves "design is done, how do I hand it to the factory?" In the traditional flow, a designer exports a STEP file, sends it to a CNC shop, and the shop's engineer responds "this R0.5 inner fillet can't be milled, minimum is R1" — sending it back. A few rounds like that and a week is gone. handoff-cad embeds manufacturer constraints into the handoff package, so you know at design time what will get rejected.
dfam-check (Design for Additive Manufacturing) is more forward-looking. It checks not "is the geometry correct" but "can it actually be printed" — overhangs beyond 45° need supports, overly thin walls warp, small holes clog after printing. This knowledge usually lives only in senior engineers' heads; now it's a skill an Agent can invoke.
---
3. Why "Skills" Instead of a "Model"
text-to-cad's key design choice: it's not a trained model, it's a set of skills.
Why does that matter? A model would need training data — millions of text→STEP pairs. That data doesn't exist. STEP files are an industrial format, not something you can scrape off the internet. Training a text-to-STEP model faces a data bottleneck an order of magnitude harder than code generation.
The skill route sidesteps this. Agents (Claude, GPT) already have spatial reasoning — they understand "a 50mm×30mm×5mm plate with M4 holes at three corners." text-to-cad gives the Agent tools: APIs for generating STEP formats, libraries for checking geometry, interfaces for querying supplier prices. The Agent handles understanding and reasoning; skills handle format conversion and domain checks.
This mirrors the architecture of addyosmani/agent-skills (general engineering skills), but in a completely different domain: physical manufacturing. General skill libraries cover code review, docs, testing; text-to-cad covers CNC, 3D printing, laser cutting, robot modeling. Complementary, not overlapping.
---
4. URDF and SDF: The Entry Point to Robot Simulation
text-to-cad isn't just about manufacturing — it covers robotics too. text-to-urdf and text-to-sdf let Agents generate robot description files.
URDF (Unified Robot Description Format) is the ROS standard, describing links, joints, and inertial parameters. SDF (Simulation Description Format) is Gazebo's format, adding sensors and environment descriptions beyond URDF.
This means you can tell an Agent "design a six-axis arm with 800mm reach and 2kg payload," get a URDF back, and drop it straight into ROS simulation. The path from "idea" to "simulation-ready" compresses into one conversation.
For robotics startups, this lowers prototyping barriers. Traditionally: mechanical engineer does CAD, controls engineer writes URDF, simulation engineer tunes Gazebo — three roles, three tools. text-to-cad automates the format-conversion work across all three; humans only make design decisions.
---
5. SendCutSend Integration: One-Click from Design to Order
The source-parts skill integrates directly with SendCutSend, an online laser-cutting/waterjet/CNC service. Give it a BOM, and it returns quotes and order links for each part.
This compresses "design → quote → order" into seconds. The traditional flow: finish design → export DXF → upload to supplier site → wait for a quote email → confirm order. With text-to-cad, the Agent sees cost in real time during design — "this part in 3mm aluminum quotes $12 at SendCutSend; switch to 2mm, $8 but not strong enough."
Front-loading cost information into the design phase is something engineering has long pursued but rarely achieved. By skill-ifying the supplier interface, text-to-cad makes the Agent a "real-time cost translator" between design and manufacturing.
---
6. Why a +75 Star Project Deserves Attention
text-to-cad gained only 75 stars today — unremarkable next to projects gaining +1000. But its lane — AI Agents × physical manufacturing — has almost no competitors.
Most AI Agent projects compete within the digital domain: better code generation, better summarization, better scraping. text-to-cad heads for the physical world. That means:
1. Different data moat. Code generation has GitHub; manufacturing has no equivalent. The skill route bypasses the data problem, but skill quality depends on the author's depth of manufacturing knowledge. 2. Different user profile. Code Agent users are programmers; CAD Agent users are mechanical engineers, robotics teams, hardware startups — a group barely covered by AI tools today. 3. Longer value chain. A code PR's chain is "write → test → merge"; a CAD file's chain is "design → inspect → quote → order → manufacture → QA." Every step is a skill opportunity.
The risks are obvious too: manufacturing has a high expertise bar, and skill correctness needs domain-expert validation. A wrong DfAM check could waste hundreds of dollars of material. But that's the nature of all professional tools — their value scales with their risk.
---
Project link: github.com/earthtojake/text-to-cad