What changed
kernelCAD has released version 0.17.0, an update to its agent-native Computer-Aided Design (CAD) system. The core philosophy of kernelCAD revolves around an "agent-first" approach, where the primary interface is not a graphical user interface (GUI) but rather editable .kcad.ts source files, a Command Line Interface (CLI), and the Model Context Protocol (MCP). This means mechanical intent is translated into deterministic source code, which then generates reviewable evidence and exportable manufacturing artifacts. The development workflow is source-centric: a design brief leads to editable .kcad.ts source, which is evaluated to produce a model. This model undergoes deterministic validation, followed by guided revision, and finally, an export package is generated. kernelCAD is not aiming to replicate traditional CAD software like Fusion or SolidWorks but focuses on providing an agent-first workflow layer above kernel technologies like OpenCASCADE and Replicad. This layer includes features such as deterministic CAD source, explicit feature history, diagnostics, validation, and variant management.
Key features of kernelCAD include a headless core that can run CAD operations in Node.js or Web Workers without requiring a DOM. It offers an Agent API for parameters, assemblies, NURBS, SDF bodies, sheet metal workflows, introspection, and feedback loops. A browser-based "Review Cockpit" allows for 3D previews of generated designs and validation results. Standard exports to STEP and STL formats are supported, built upon a robust kernel, with OpenCASCADE as a primary dependency and Replicad utilized where applicable.
For agents and contributors, the .kcad.ts source file is the definitive source of truth for designs. Rendered PNG/MP4 files, STEP/STL exports, score JSON, and capture-run metadata are considered generated evidence or deliverables. Changes should be made to the source code first, followed by regenerating explicit targets. Deterministic checks are encouraged before visual judgment, using commands like kernelcad evaluate model.kcad.ts and kernelcad export step model.kcad.ts. For visual evidence, a deterministic inspection bundle can be generated using kernelcad render inspect model.kcad.ts. This bundle can include machine-readable channels such as mask, depth, and normals in addition to RGB views.
Why it matters for builders
This release is significant for AI builders and developers as it fundamentally redefines the interaction model for CAD. By treating geometry as deterministic source code, kernelCAD opens up new possibilities for AI agents to generate, manipulate, and validate complex designs programmatically. The emphasis on an editable source-first workflow means that AI models can directly interact with and modify design parameters, enabling more sophisticated automated design processes. Furthermore, the built-in validation and review mechanisms provide a structured way to ensure design integrity, which is crucial for AI-generated outputs.
Practical impact
Developers can get started by installing kernelCAD globally via npm: npm install -g kernelcad. A simple example involves creating a bracket.kcad.ts file with parameters for width, height, and thickness, defining a basic box shape with a subtracted cylinder and fillet. This script can then be evaluated and exported to STL using CLI commands. For integration with AI agents, kernelCAD functions as an MCP server. Users can configure Claude Desktop to connect to a local kernelCAD MCP server by adding an entry to its configuration file. This integration allows Claude Desktop to utilize kernelCAD's tools for design review, model introspection, and export. For prompt-driven authoring, the hosted gateway at app.kernelcad.com/connect offers a generate_kcad_from_prompt function, translating natural language briefs into .kcad.ts source code, alongside introspection and review tools. A free tier provides 3 LLM-generated parts per month, with unlimited access to introspection and review tools.
Caveats and source limits
The source indicates that kernelCAD is built on OpenCASCADE and Replicad, but the specific versioning or details of these underlying kernels are not provided. While the release is marked as "fresh release" and version v0.17.0 is specified, there are no independent benchmark results or performance metrics detailed in the provided excerpts. The integration with Claude Desktop is described, but the exact capabilities and limitations of the MCP server interaction are not fully elaborated. Attribution details mention that exported files may name kernelCAD and some viewer pages show a "Made with kernelCAD" link, with potential referral links for vendor geometry, but the specifics of self-hosting or fork modifications for attribution are not deeply detailed.
Sources
Claim check: 10/10 supported claims - 10 evidence links - 100% avg confidence
- kernelCAD turns mechanical intent into editable `.kcad.ts` source, deterministic review evidence, and exportable manufacturing artifacts.supported - github.com
- kernelCAD is an agent-first workflow layer above kernel technologies like Replicad and OpenCASCADE.supported - github.com
- kernelCAD offers deterministic CAD source, feature history, diagnostics, validation, and variants.supported - github.com
- kernelCAD can run CAD operations in Node.js or Web Workers without a DOM.supported - github.com
- kernelCAD provides an Agent API for parameters, assemblies, NURBS, SDF bodies, sheet metal workflows, introspection, and feedback loops.supported - github.com
- kernelCAD supports standard exports to STEP and STL formats.supported - github.com
- kernelCAD can be used with Claude Desktop via MCP server integration.supported - github.com
- The hosted gateway at app.kernelcad.com/connect offers a `generate_kcad_from_prompt` function.supported - github.com
- The hosted gateway offers a free tier with 3 LLM-generated parts per month and unlimited introspection/review tools.supported - github.com
- kernelCAD version v0.17.0 has been released.supported - github.com
Caveats
- Single-source caution: verify critical details at the linked source.
Radar score 88/100 - how it was calculated
- Reliability 82: GitHub metadata supports source trust
- Freshness 92: Fresh GitHub activity
- Novelty 77: Fresh GitHub release
- Technical 89: Repository technical metadata
- Developer 96: Developer tooling signals
- Ecosystem 72: Fresh GitHub release
- Confidence 100: Claims have reliable evidence