text-to-cad: A Skill Library for Turning Plain Language into CAD and Robot Files

What text-to-cad Actually Solves

text-to-cad is not a CAD application. It is a library of agent skills that let an LLM-based coding agent create, inspect, and hand off hardware artifacts. The README describes it as a collection for generating, inspecting, sourcing, slicing, and handing off CAD and robot-description files from local project files. The target user is an engineer who already works with agents like Codex or Claude Code and wants to extend them into mechanical design. Instead of switching between a chat tool and a CAD program, you ask the agent to produce a STEP file, a URDF, or a G-code slice, and the skill provides the structured workflow. This matters because generic agents often hallucinate file formats or omit required fields. The skills encode the steps, so the output is more likely to be valid.

How the Skills Are Organized and What Each One Covers

The repository lists twelve skills in a table, each with a summary and a link to its SKILL.md. The CAD skill creates and edits models from text or images, with STEP as the primary output and STL, 3MF, and GLB as export options. The URDF skill writes robot structure files with links, joints, limits, inertials, and meshes. SRDF adds MoveIt planning groups, end effectors, poses, and collision rules to a URDF. SDF builds simulator models and worlds with frames, physics, sensors, and lights. There are also skills for DXF 2D drawings, finding off-the-shelf STEP parts, checking files for SendCutSend, measuring mesh printability (DfAM Check), slicing to G-code with real slicer CLIs, and controlling Bambu Lab printers. The Implicit CAD skill is marked experimental and uses GLSL signed-distance fields with raymarch rendering in a browser. This is a broad spread, but each skill is a focused workflow, not a generic prompt.

Installation: The Skills CLI and Provider Plugins

The preferred installation path is the Skills CLI. The README gives this command: npx skills add earthtojake/text-to-cad. This installs the individual skills directly for supported agents. The same command updates the package, because add re-fetches and overwrites existing skills, and it also picks up any new skills added in newer releases. The README warns that npx skills update only refreshes skills already in your lockfile, so it silently misses new ones. That is a real operational detail: if you rely on update, you may not get newly added skills. To remove a retired skill, you use npx skills remove <skill>. There is also an undocumented alias, npx skills install, which works the same as add. For provider-native installs, the README shows commands for Codex, Claude Code, and Grok Build. Codex requires version 0.142.0 or newer, and on older versions the plugin is skipped silently, so you must upgrade with npm install -g @openai/codex@latest.

The Plugin Marketplace Mechanism and Its Version Trap

The plugin installation uses a marketplace file. For Codex, you run codex plugin marketplace add earthtojake/text-to-cad and then codex plugin add cad@text-to-cad. Claude Code uses claude plugin marketplace add and claude plugin install. Grok Build reuses the existing .claude-plugin/marketplace.json and has no separate manifest. The version trap is explicit: Codex resolves the repository-root plugin only from 0.142.0 onward. On older versions, the plugin is skipped silently and never appears in codex plugin list. That is a failure mode you need to plan around. If you are on an older Codex, you will not see an error; the plugin just will not exist. The README tells you to restart your agent if newly installed skills do not appear. This is a concrete operational constraint that affects real adoption.

Development Workflow and Branching Strategy

Development happens on the develop branch, and pull requests should target develop, not main. The README says that main contains the generated skill outputs needed by provider installers. This is an important distinction: main is a release branch with pre-built artifacts, while develop is for active work. For local contribution, the README points to CONTRIBUTING.md for a symlink workflow and skill linking. This suggests that the skills themselves are not just markdown files; they may reference generated outputs or dependencies. The README mentions skills/cad/requirements.txt, which implies Python dependencies for the CAD skill. This is a practical consideration: installing text-to-cad may pull in more than just JavaScript, depending on the skill. The repository's primary language is JavaScript, but the CAD skill has a requirements file, so you may need a Python environment for that skill to work.

Limitations and Failure Modes: When text-to-cad Is the Wrong Tool

The most obvious limitation is that text-to-cad is a set of agent skills, not a standalone application. It requires an agent that supports the Skills CLI or the plugin system. If you are not using Codex, Claude Code, or Grok Build, you may have to rely on the CLI, which may not integrate with your agent. The README does not list a full set of supported agents, so you need to verify compatibility. Another limitation is the experimental nature of Implicit CAD, which uses GLSL signed-distance fields. That is not a mature path for production parts. Also, the G-code skill relies on real slicer CLIs, so you need those installed and configured. The DfAM Check skill measures printability, but it does not guarantee that a model is printable; it only provides measurements. Finally, the update mechanism has a trap: npx skills update misses new skills, so you must remember to use add to get the latest. This is a maintenance cost that can lead to missing features if you are not careful.

Alternatives: How Other Approaches Differ

A direct alternative is using a dedicated CAD automation library like CadQuery or OpenSCAD, where you write parametric scripts in Python or a custom language. The difference is that text-to-cad is agent-driven: the LLM writes the code or commands, and the skill provides the workflow. With CadQuery, you write the script yourself, which gives you full control but requires programming knowledge. Another alternative is using a cloud CAD API like Autodesk Platform Services, which offers REST endpoints for model translation and viewing. That approach is more heavyweight and requires API keys and a server, whereas text-to-cad runs locally. A third alternative is to use a general-purpose agent and prompt it with your own CAD instructions, but that lacks the structured validation that text-to-cad's skills provide. The key difference is that text-to-cad encodes domain-specific steps, so the agent is less likely to produce malformed files.

Maintenance, Licensing, and Upgrade Cost

The project is licensed under MIT, which means you can use, modify, and distribute it freely, with attribution. That is a permissive license, but it does not come with warranty, so you are responsible for verifying the output. The repository is actively maintained, with releases on consecutive days (0.4.28, 0.4.27, 0.4.26). The README notes that releases add skills, so the upgrade cost is real: you need to run npx skills add to get new skills, and you must check the release notes for breaking changes. The development branch is develop, and main is for generated outputs, so if you contribute, you need to follow the branching workflow. The maintenance cost is moderate: you need to keep the skills updated, and you need to handle the fact that retired skills are not removed automatically. You also need to ensure your agent version matches the plugin requirements, as seen with Codex 0.142.0.

Editorial conclusion

Adopt text-to-cad if you are an engineer or maker who wants to generate CAD, robot description files, or G-code directly from natural language inside an agent like Codex or Claude Code, and you can tolerate the experimental edge of implicit CAD. Do not adopt it if you need a fully automated, production-grade CAD pipeline with no manual review, or if you rely on stable releases without frequent updates. Before adopting, verify that your agent version meets the minimum requirements (Codex 0.142.0 or newer for the plugin) and test the CAD skill on a simple part to check the STEP output quality. Then confirm that the skills you need exist in the current release, since new skills are added regularly and old ones are not removed automatically.

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes