kernelCAD
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@kernelCADevaluate bracket.kcad.ts and export step"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
kernelCAD - CAD runtime for agents
kernelCAD turns mechanical intent into editable .kcad.ts source,
deterministic review evidence, and exportable manufacturing artifacts.
The core loop is source-first: design brief -> editable .kcad.ts ->
evaluated model -> deterministic validation -> guided revision -> export package.
kernelCAD is not trying to be a browser clone of Fusion, Onshape, or SolidWorks. Replicad/OpenCASCADE are the kernel layer. kernelCAD is the agent-first workflow layer above the kernel: deterministic CAD source, feature history, diagnostics, validation, variants, and human review.
Install
Already in ChatGPT and Claude: no install
kernelCAD runs as a connected app in ChatGPT (Apps SDK, with the Studio viewer and paint review inside the chat) and in Claude / claude.ai (custom connector). Add it from the chat settings and sign in:
ChatGPT: connect guide or ChatGPT → Settings → Connectors
Claude: connect guide or claude.ai → Settings → Connectors, URL
https://mcp.kernelcad.com/mcp
One account everywhere. The coding-agent plugins below use the same hosted
server (https://mcp.kernelcad.com/mcp) and the same kernelCAD sign-in. A model
saved from Claude Code, Codex, Cursor, Gemini or Grok opens at the same Studio
share link (/p/<slug>) you see in ChatGPT or Claude. Sign in with the same
account in each client.
Ask your agent: “look at kernelcad.com and install the mcp”. Agent-readable installation instructions cover hosted setup, local npm setup, and connection verification.
Coding agents: one command
One command connects your agent to the hosted kernelCAD MCP server
(https://mcp.kernelcad.com/mcp) and adds the kernelCAD skills. Nothing runs
locally and no OpenCASCADE install is needed. Anonymous use works; sign in
through your client's OAuth prompt for account features.
Agent | One-liner |
Claude Code |
|
Codex |
|
Cursor | Add kernelCAD MCP to Cursor (MCP only) |
Gemini CLI |
|
Grok Build |
|
VS Code / Copilot |
|
Skills only (any agent) |
|
Inside Claude Code you can also type /plugin marketplace add w1ne/kernelCAD-web
then /plugin install kernelcad@kernelcad. Restart the app after installing.
Claude Desktop, ChatGPT, Windsurf, Cline, Zed, Continue, update, uninstall and the Cursor skills install are in
docs/usage/agent-plugin-install.md.
Related MCP server: gitcad
Showcase
Each model below started from the plain-language prompt shown. An AI agent wrote it through kernelCAD's own authoring loop (CLI / MCP): it wrote .kcad.ts source, ran evaluate, rendered the seven-view inspection sheet, read the renders, fixed what was wrong, and repeated. None of these is a one-shot generation. Each folder in showcase/ holds the parametric source, the turntable GIF (kernelcad animate --turntable, publish look), and a short note on what to look at and which checks were run.
Model | Prompt | Params | Preview |
Hollow shell ~150 mm tall, tapers 70 → 50 mm, hex section twisting 120°, 2 mm wall, closed bottom |
| ||
Ø40 × 22 mm, diamond-knurled grip, raised pointer, 6 mm D-shaft bore, M3 set-screw hole |
| ||
Filleted L-bracket with a generative hex lattice |
| ||
True airfoil from the NACA equations, tapered loft |
| ||
Meshing bevel gear pair at 90° |
| ||
7 swept backward-curved blades |
| ||
Sun, planets, ring, carrier |
| ||
Round crankcase, nine finned cylinders in a star, domed heads, prop hub |
|
Some checks go past the picture:
Mechanisms that are checked, not just drawn. The planetary stage and the bevel drive are assemblies with revolute mates, gear ratios coupled from the tooth counts, and declared gear-pair transmissions. kernelCAD's mechanism-truth check (
evaluate_script→mechanism: "real") turns each mechanism through its limits with the couplings applied and looks for overlapping teeth.kernelcad validate --include-interferencereports both assemblies clean.Manufacturing evidence. The honeycomb bracket declares an FDM print gate.
kernelcad dfmpasses it (thinnest wall 2.0 mm, no unsupported overhang), andkernelcad export stepwrites a B-rep whose holes read back withkernelcad inspect step.
Philosophy
Agent-First: The primary interface is
.kcad.ts, CLI, and MCP. The UI reviews and steers generated designs.Editable Source: Geometry is deterministic source code with explicit feature history, params, assemblies, and diagnostics.
Validation-Centered: Useful CAD output needs deterministic review for constraints, hardware fit, clearances, and exportability.
Workflow Over Primitives: Product progress is measured by intent-level workflows, not by copying every traditional CAD button.
Agent workflow
For agents and contributors, .kcad.ts source is the design source of truth.
Rendered PNG/MP4 files, STEP/STL exports, score JSON, and capture-run metadata
are generated evidence or deliverables. Prompt briefs, source files, and
provenance metadata remain source. Change source first, regenerate explicit
targets, and avoid broad artifact refreshes unless you are intentionally
rebuilding a demo or release bundle.
Use deterministic checks before visual judgment:
kernelcad evaluate model.kcad.ts
# then run review_cad from an MCP client connected through `kernelcad mcp`
kernelcad export step model.kcad.ts -o model.step
kernelcad export stl model.kcad.ts -o model.stlWhen a rendered artifact is produced, inspect the image/video itself and report what is visible: proportions, interfaces, floating geometry, occlusion, camera framing, and whether required features are legible. Passing tests are not a substitute for visual evidence.
When visual evidence matters, keep a deterministic inspection bundle:
kernelcad render inspect model.kcad.ts model.inspectThe v1 bundle writes a manifest and canonical RGB views. It can also emit
machine-readable mask, depth, and normals channels:
kernelcad render inspect model.kcad.ts model.inspect --channels rgb,mask,depth,normalsUse --focus <names> or --hide <names> when a specific feature id or
assembly part needs isolated review. Keep richer channels in the same manifest
packet; do not replace the canonical RGB views.
For real hardware, prefer catalog/vendor geometry via lib.fromSTEP(...)
instead of fake placeholder boxes or cylinders. This keeps mounts, clearances,
and mechanism checks tied to physical components.
Features
Headless Core: Run CAD operations in Node.js or Web Workers without a DOM.
Agent API: kernelCAD APIs for params, assemblies, NURBS, SDF bodies, sheet metal workflows, introspection, and feedback loops.
Review Cockpit: Browser-based 3D preview for inspecting generated designs and validation results.
Standard Exports: STEP/STL generation.
Robust Kernel: Built on OpenCASCADE, with Replicad kept as a kernel implementation detail where it is still used.
Get Started
npm install -g kernelcadDrop this in bracket.kcad.ts:
const w = param('Width', 60, { unit: 'mm' });
const h = param('Height', 40, { unit: 'mm' });
const t = param('Thickness', 5, { unit: 'mm' });
const base = box(w, h, t);
const hole = cylinder(t + 2, 4).translate(w / 2, h / 2, -1);
return base.subtract(hole).fillet(1);Run it:
kernelcad evaluate bracket.kcad.ts
kernelcad export stl bracket.kcad.ts -o bracket.stlThat's it. For agents: kernelcad mcp runs an MCP server with dynamic model
introspection tools. Object tool responses include structuredContent alongside
the serialized JSON text so MCP clients can validate advertised output schemas.
See SKILL.md (bundled with the install) for the full API
surface and authoring guide.
Use with Claude Desktop
kernelCAD runs as an MCP server, so Claude Desktop can drive it locally. After
npm install -g kernelcad, add an entry to your Claude Desktop config file:
macOS:
~/Library/Application Support/Claude/claude_desktop_config.jsonWindows:
%APPDATA%\Claude\claude_desktop_config.jsonLinux:
~/.config/Claude/claude_desktop_config.json
{
"mcpServers": {
"kernelcad": {
"command": "npx",
"args": ["-y", "kernelcad", "mcp"]
}
}
}Restart Claude Desktop. The kernelCAD MCP tools (review_cad, model
introspection, export helpers, etc.) appear in the tools list.
The first full local evaluate_script build returns viewerUrl. Open it in
Studio once: later full builds and source-edit tool calls replace the source,
history, and geometry behind that same session link. The viewer is served on
127.0.0.1 by the MCP process and lasts until that process closes; it is not a
public share link. Dry runs leave it unchanged. A build failure keeps the last
built model visible; a successfully built model with DFM failures remains
viewable. Build diagnostics shown there are not a full CAD/physics review.
Published packages include the Studio bundle. When running MCP from a checkout,
run npm run build:mcp-viewer first. If the bundle cannot be served, the build
result includes viewerError separately from its CAD diagnostics. Hosted MCP
and --cloud continue to use the gateway's Studio/project URLs.
For hosted workflows, sign in at the existing connection guide. It gives the current connector and OAuth steps for your client. The hosted tool set depends on authentication, your selected profile, and server-native tools; use the connected server's discovery rather than assuming the local list.
The local installer is available. Preview the config change before applying it:
npx -y kernelcad install --claude-desktop --dry-run
npx -y kernelcad install --claude-desktopOther hosted setup guides: ChatGPT, Claude Code, Codex, Cursor, and VS Code.
Browse the generated local MCP tool reference for complete input/output schemas and behavior annotations, or search the design guide for existing recipes and cited, dated manufacturing catalog snapshots.
For local Studio, get_selection and set_selection
share face/edge picks with your MCP agent. Open the same script in Studio and
use its session token; picks expire on model rebuild.
For opt-in kernel selection and fallback diagnostics, see the experimental guarded SMLib kernel guide and its workflow demo bundle.
Contributing
Web app + dev workflow (clone the repo, npm install, npm run dev) for contributors who want to hack on the kernel or the visual debugger:
git clone https://github.com/w1ne/kernelCAD-web.git
cd kernelCAD-web
npm install
npm run dev # web visual debugger at localhost:5173
npm run build:cli # build the CLI bundle into dist/cli/
npm run test # full vitest suite
npm run qc # quick quality gate (lint + typecheck + tests)Attribution and referral links
Exported files name kernelCAD in their metadata field (STEP FILE_NAME, GLB
asset.generator, the STL header, a DXF or XML comment), and public viewer
pages show a small "Made with kernelCAD" link. Some tools return vendor links
that can carry a referral tag; results with a tagged link include the
disclosure "Some links are referral links; kernelCAD may earn a commission."
A vendor with no tag gets the clean URL. Self-hosters and forks change
all of it in two files, or with KERNELCAD_REFERRAL_<VENDOR> and
KERNELCAD_REFERRALS=off. See ATTRIBUTION.md.
License
MIT — see LICENSE. The kernelCAD name is covered by the trademark policy.
This server cannot be deployed
Maintenance
Related MCP Connectors
- OwlCADOAuthcom.owlcad
Parametric 3D CAD for AI agents: build print-ready parts, check them, export STL, 3MF or STEP.
Design domain models and generate deterministic multi-stack code, driven by your coding agent.
Headless CAD engine & MCP server for 2D/3D drafting, solids, DXF/SVG export, and automation.
OCR, transcription, file extraction, and image generation for AI agents via MCP.
Related MCP Servers
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to write TypeScript to create, render, verify, and export parametric 3D models using Replicad/OpenCascade CAD kernel, headlessly from terminal or MCP clients.6MIT

gitcadofficial
AlicenseNot gradedqualityBmaintenanceEnables agents to create and manipulate 3D CAD models using natural language via MCP, with git-based versioning and manufacturing output generation.1Apache 2.0- AlicenseNot gradedqualityBmaintenanceEnables engineering agents to perform parametric CAD operations through a controlled MCP gateway built on FreeCAD, with transactions, diagnostics, and reproducible verification.2GNU Lesser General Public v2.1 only
- FlicenseNot gradedqualityBmaintenanceEnables agent-assisted CAD engineering, allowing users to create, validate, and export CAD designs through natural language, with a deterministic engine that has zero LLM runtime dependency.-