Open Source Text to CAD Harness: A Deep Dive

Demo image: Demo of the text-to-cad harness generating and previewing CAD geometry

In a world where 3D models are increasingly authored by intelligent coding agents, the Open Source Text to CAD Harness stands out as a transparent, extensible platform for generating, inspecting, and exporting CAD geometry entirely from source-controlled, programmable workflows. This harness is designed to empower developers, researchers, and makers to describe parts, assemblies, fixtures, or robots in code, then iteratively refine, reproduce, and export precise CAD outputs without ever leaving their favorite coding environment.

The project embraces openness, offering bundled skills for CAD creation, robot description, motion planning, and manufacturing preflight checks. It ships with ready-to-use components and clear pathways for extension, enabling teams to tailor the workflow to their specific design and manufacturing pipelines.

What is the Text to CAD Harness?

At its core, the harness provides a conduit between naturalistic or code-driven design prompts and fully realized CAD artifacts. It enables you to:

• Describe a part, assembly, or mechanism in text or code, then translate that description into source-controlled CAD sources.
• Generate and export exact geometric targets in common formats such as STEP, STL, 3MF, DXF, GLB, and URDF, among others.
• Browse and inspect generated geometry inside a CAD Explorer-like environment to verify fit, tolerances, and visual fidelity.
• Reference stable handles like @cad[...] to enable geometry-aware follow-up edits by agents or teammates.
• Render quick visual previews during iteration loops to accelerate decision-making.
• Reproduce work by editing the source first and regenerating explicit targets—promoting traceability and collaboration.
• Run locally without backend hosting for a lightweight, offline workflow.

The harness is also designed with a practical, project-focused workflow in mind. It serves as a launchpad for autonomous design loops while keeping CAD artifacts aligned with the original source code.

Why this Harness Stands Out

• Open Source and Transparent: The project is MIT-licensed and hosted with a transparent build and test story, inviting contributions from a broad community.
• Bundled Skills for CAD and Robotics: The harness ships with targeted capabilities for CAD generation, robotic descriptions, motion planning, and manufacturing preflight checks, all living under a single, canonical location.
• Reproducible Workflows: By keeping links to stable references and explicit target generation, teams can reproduce results reliably across machines and over time.
• Local-First Architecture: The system is designed to run locally, minimizing dependencies on centralized backends and enabling offline development and testing.
• Rich Export Options: From traditional CAD formats to robot descriptions and motion data, the export surface covers a wide range of downstream workflows.

The project also weaves together a constellation of skills and repositories that expand the harness’s utility beyond a single project. The CAD Skill, URDF Skill, and Robot Motion Skill each provide specialized tooling, while a set of symlinks and canonical locations ensures compatibility with Claude Code and other agents.

🧰 Bundled Skills: A Closer Look

The harness brings a curated set of file-targeted skills that are essential for end-to-end CAD-driven design and robotics workflows. These bundled skills can be used as-is for local CAD projects, or as a template for installing and integrating independent skill repositories.

• CAD Skill: This skill handles STEP, STL, 3MF, DXF, GLB/topology exports, image rendering, and geometry references via @cad[...]. It’s designed to produce stable geometry-ready outputs and to be referenced by other agents during follow-up edits.

• Bundled skill page: .agents/skills/cad/SKILL.md

• Standalone repository: https://github.com/earthtojake/cad-skill

• URDF Skill: Generates URDF XML with robot links, joints, limits, validation, and mesh references. It’s a critical bridge between CAD geometry and robot simulations or ROS ecosystems.

• Bundled skill page: .agents/skills/urdf/SKILL.md

• Standalone repository: https://github.com/earthtojake/urdf-skill

• Robot Motion Skill: Focused on ROS 2/MoveIt integration, CAD Explorer motion artifacts, inverse kinematics, path planning, and motion-server testing for existing URDFs. It enables realistic motion planning around CAD-generated geometries.

• Bundled skill page: .agents/skills/robot-motion/SKILL.md

In addition, the skills live canonically under a single directory, .agents/skills, which keeps core capabilities organized and discoverable. Claude Code compatibility is provided by per-skill symlinks under .claude/skills, ensuring agents with Claude Code can leverage the same skill set.

📸 Screenshots and Demos

Visual feedback is an essential part of designing with CAD from code. The project ships with a set of dynamic demonstrations that illustrate how the harness operates in real-time.

• CAD Skill demo: See the generated geometry in CAD Explorer.
  Reference: .assets/text-to-cad-demo.gif

• URDF Skill demo: Observe how robot descriptions are produced and inspected in CAD Explorer.
  Reference: .assets/urdf-demo.gif

• Robot Motion Skill demo: Watch inverse kinematics and motion planning applied to an existing URDF.
  Reference: .assets/robot-motion-demo.gif

These visuals give you a window into the end-to-end loop: describe, generate, inspect, and refine geometry and robot models with a few clicks or lines of code.

🔁 The Core Workflow

The harness is designed to support a repeatable, six-step workflow that you can adopt for any CAD-centric project:

1) Describe: Tell your preferred coding agent about the part, assembly, fixture, robot, or mechanism you want. The goal is to capture the intent clearly so the agent can translate it into precise CAD sources.

2) Edit: Let your coding agent update repository-local CAD source files. The edits can be incremental, targeted, and aligned with the project’s source control.

3) Regenerate: Create explicit targets in formats like STEP, STL, 3MF, DXF, GLB, or URDF to lock in the design.

4) Inspect: Open CAD Explorer to review the generated model, verify geometry, and validate design choices against requirements.

5) Reference: Copy stable @cad[...] handles for geometry-aware edits in subsequent steps, ensuring continuity and precision.

6) Commit: Save both the source and the generated artifacts together, preserving the relationship between intentions, code, and CAD output.

This workflow makes it feasible to iterate quickly while keeping design provenance intact—a critical advantage in collaborative engineering environments.

🚀 Benchmarks and Test Suite

The repository includes a set of benchmark scenarios designed to exercise common CAD tasks. To optimize for lightweight clones, benchmark assets are stored under .assets/** and are Git LFS–managed. Local hydration of assets can be done selectively to avoid pulling large assets during standard clones.

Here are the ten benchmark targets, described succinctly, each paired with its own output visualization:

• 1) Rectangular calibration block with four holes

• Prompt: Create a centered 100 x 60 x 20 mm block with four 8 mm vertical through-holes. Add only a 2 mm chamfer on the top outer perimeter.

• Output: Rectangular calibration block orbit gif

• Visualization: .assets/benchmarks/benchmark01rectangularcalibrationblock.gif

• 2) Circular flange with bolt-hole pattern

• Prompt: Create an 80 mm diameter, 10 mm thick circular flange with a 30 mm central through-bore. Add six 6 mm through-holes on a 60 mm bolt circle and fillet the outside circular edges.

• Output: Circular flange orbit gif

• Visualization: .assets/benchmarks/benchmark02circular_flange.gif

• 3) L-bracket with gussets and two hole directions

• Prompt: Create an L-bracket from a base plate and rear vertical plate. Add vertical base holes, horizontal back-plate holes, two triangular gussets, and a filleted base/back transition.

• Output: L-bracket orbit gif

• Visualization: .assets/benchmarks/benchmark03l_bracket.gif

• 4) Stepped shaft with keyway

• Prompt: Create a 120 mm shaft along X with 20/30/20 mm diameter stepped sections. Add end chamfers and a shallow rectangular keyway on top of the middle section.

• Output: Stepped shaft orbit gif

• Visualization: .assets/benchmarks/benchmark04steppedshaftkeyway.gif

• 5) Open-top electronics enclosure with bosses

• Prompt: Create a hollow open-top enclosure with 3 mm walls and floor. Add four internal standoffs with centered blind holes and 2 mm outside vertical corner fillets.

• Output: Open-top electronics enclosure orbit gif

• Visualization: .assets/benchmarks/benchmark05opentopelectronics_enclosure.gif

• 6) Aerospace-style clevis bracket with lightening cutouts

• Prompt: Create a symmetric clevis bracket with a base plate, two rounded lugs, base mounting holes, and a horizontal lug bore. Add triangular lightening cutouts, reinforcing ribs, and rounded transitions.

• Output: Clevis bracket orbit gif

• Visualization: .assets/benchmarks/benchmark06clevisbracketlightening_cutouts.gif

• 7) Radial-engine-style cylinder with cooling fins

• Prompt: Create a vertical engine-cylinder form with a central barrel, 12 cooling fins, a base flange, and a top cap. Add a 35 degree angled spark-plug boss with a coaxial through-hole.

• Output: Radial-engine-style cylinder orbit gif

• Visualization: .assets/benchmarks/benchmark07radialenginecylinder.gif

• 8) Centrifugal impeller with backward-curved blades

• Prompt: Create a centrifugal impeller with a backplate, hub, and through-bore. Add 12 fused backward-curved blades sweeping about 45 degrees from root to tip.

• Output: Centrifugal impeller orbit gif

• Visualization: .assets/benchmarks/benchmark08centrifugal_impeller.gif

• 9) Spiral staircase with helical handrail

• Prompt: Create a miniature spiral staircase with a central column, base disk, and 20 rising wedge treads. Add a one-revolution helical handrail and vertical balusters at the tread outer ends.

• Output: Spiral staircase orbit gif

• Visualization: .assets/benchmarks/benchmark09spiral_staircase.gif

• 10) Simplified planetary gear stage

• Prompt: Create a flat planetary gear assembly with separate sun, planet, ring, carrier, and pin bodies. Use simplified trapezoidal teeth and place three planets around the sun on a 42 mm radius circle.

• Output: Planetary gear stage orbit gif

• Visualization: .assets/benchmarks/benchmark10planetarygearstage.gif

These benchmarks not only exercise geometry generation but also encourage robust workflows for validation, visualization, and reproducibility. By documenting each scenario along with its corresponding animation, the project provides an accessible way to understand how the text-to-CAD pipeline behaves across a spectrum of design challenges.

A quick note on how benchmarks are hydrated locally: to fetch only the benchmark assets, you can run a targeted Git LFS pull, which helps keep lightweight clones snappy while still enabling full experimentation when needed.

🧭 Quick Start: Getting Up and Running

For developers who want to jump in quickly, the project offers a straightforward path to clone, set up, and start running the harness with local CAD Explorer support. The quick start is designed for reproducibility and minimizes the friction of bringing up a full CAD environment.

• Clone the repo and enter the project directory:

• git clone https://github.com/earthtojake/text-to-cad.git

• cd text-to-cad

• Install Python CAD dependencies (using a Python 3.11+ virtual environment):

• python3.11 -m venv .venv

• ./.venv/bin/python -m pip install --upgrade pip

• ./.venv/bin/pip install -r .agents/skills/cad/requirements.txt

• Optional: install other bundled skill requirements if you need URDF or robot workflows:

• ./.venv/bin/pip install -r .agents/skills/urdf/requirements.txt

• Install CAD Explorer dependencies (Node.js):

• npm --prefix .agents/skills/cad/explorer install

• Run the local CAD Explorer:

• npm --prefix .agents/skills/cad/explorer run dev

• Then open http://localhost:4178

• For root-aware agent workflows across multiple projects, you can reuse a matching server or start one on a free port:

• npm --prefix .agents/skills/cad/explorer run dev:ensure -- --file STEP/sample_part.step

• Then open the URL printed by the command.

This quick-start path emphasizes the offline, local-first nature of the harness while still supporting the broader ecosystem of CAD exploration and automation.

🧰 Practical Implementation Notes

• The canonical skills live under .agents/skills, which keeps the architecture clean and predictable for local development.
• Claude Code compatibility is facilitated by per-skill symlinks under .claude/skills, ensuring that different AI agents can engage with the same skill set without friction.
• Export targets are diverse and include standard formats that are widely used in industry and academia—STEP, STL, 3MF, DXF, GLB—and specialized formats such as URDF for robotics.

For teams building automated design pipelines, these details matter: stable export formats enable downstream simulation, manufacturing, and integration with robotics workflows, while geometry references via @cad[…] ensure that subsequent edits remain precise and traceable.

📑 Reference and Reproducibility

The harness is designed to help teams reproduce designs across iterations. By storing both the CAD sources and the explicit target artifacts in version control, teams can clearly track how a design evolved from concept to a production-ready geometry. The reference handles, such as the stable @cad[...] syntax, act as anchors in the design space, making it easier for agents to apply geometry-aware edits in a predictable way.

The bundling approach also encourages a stable baseline for projects. Developers can rely on the bundled SKILL.md documentation within each skill directory to understand expected inputs, outputs, and validation steps. When needed, standalone repositories offer a path to decoupled workstreams, making it straightforward to integrate new tools or swap out components in a larger automation pipeline.

🌍 Community, Licensing, and Collaboration

The project embraces an inclusive, open collaboration model. The MIT license invites broad usage, experimentation, and adaptation, while GitHub hosting provides visibility, issue tracking, and contribution workflows. The badge ecosystem you might see in the original repository—covering stars, forks, and platform integrations—reflects a community-driven effort to share knowledge and tooling for text-to-CAD workflows.

As an open-source project, it invites engineers to contribute in ways that align with their needs—whether that means adding new export formats, extending robot motion capabilities, or integrating with additional CAD environments. The fusion of CAD generation, robot description, motion planning, and manufacturing readiness in a single harness is designed to accelerate collaborative design while preserving a clear, auditable lineage from prompt to production artifact.

💡 Use Cases and Scenarios

• A design engineer uses a Codex- or Claude-powered agent to describe a new bracket, then iterates with CAD Explorer previews to validate fit and strength.
• A robotics researcher defines a new manipulator in URDF, exports the geometry, then runs a motion plan through a MoveIt-based workflow to verify reachability and collision avoidance.
• A manufacturing engineer generates an enclosure with internal standoffs and fillets, exports a STEP file for tooling and inspection, and stores the design for supplier review.
• An educator demonstrates the end-to-end loop of describing a mechanism in code, generating CAD, and validating through quick visual checks—perfect for classroom learning and hands-on labs.

📝 Concluding Thoughts

The Open Source Text to CAD Harness is more than a collection of tools; it’s a philosophy for modern design automation. By uniting code-driven CAD generation with robust export options, accessible previews, and a clear path for reproducible workflows, it lowers the barrier to entry for teams exploring automated design and robotics pipelines. The bundled skills provide a solid foundation for common workflows, while the architecture encourages extension and experimentation, enabling researchers and practitioners to push the envelope in CAD-enabled automation.

If you’re exploring how to harness coding agents to generate, refine, and export precise CAD geometry, this harness offers a compelling, well-documented starting point. With local-first operation, a transparent skill ecosystem, and a rich set of export targets, it supports iterative design, rigorous validation, and productive collaboration across disciplines.

[Screenshots and demos continue to illustrate the journey, with CAD, URDF, and Robot Motion perspectives to guide your own experiments and implementations.]

• CAD Art and Preview:
• URDF Robot Description:
• Robot Motion Feedback:

In short, the Text to CAD Harness helps you describe, generate, inspect, and perfect CAD models—and do so in a way that embraces collaboration, reproducibility, and extensibility. Whether you’re a developer, a researcher, or a practitioner preparing designs for manufacturing or robotics, this platform offers a robust, open pathway to bring code-driven CAD from concept to production-ready artifacts.