Authoring CAD with AI

Ask for a part in plain English. The agent writes it, runs it on a real CAD kernel, and proves it is correct before handing it back.

   "a 40×20×10 mm bracket with two M4 holes"
                  │  agent writes bracket.brep.ts and runs
                  ▼  it on a real OpenCascade kernel
   ────────────────────────────────────────────────────
   ✓ valid solid       ✓ volume ≈ 7,750 mm³ (within 1%)
   ✓ snapshots rendered    →  bracket.step ready to hand off

An LLM can't see geometry. Code that reads correctly can still produce a solid that is broken, empty, or simply the wrong size — and the model has no way to notice. brepjs-verify takes the guesswork out: the agent writes the part, runs it against a real kernel, and reads back what the kernel actually measured instead of judging the code by eye.

What it checks

Every part gets three independent verdicts, so a pass means more than "the code ran":

• Validity — is it a real manifold solid with positive volume? (the kernel's validSolid brand)
• Intent — does it match the dimensions you asked for? (measureVolume / measureArea / bounds vs an expected block)
• Shape — does it actually look like the request? (multi-view PNG snapshots, for the agent or you to review)

STEP is the validated deliverable; GLB, STL, and snapshots are derived previews.

What you install

brepjs-verify is two parts that work together:

• The skill teaches the agent the workflow — brief → author → verify → repair — and carries the API references and examples it learns from.
• The runtime is the CLI the skill drives. It loads your part on a real kernel, checks validity, measures volume / area / bounds, renders snapshots, and exports STEP.

They install separately because Claude Code loads skills from plugins (a git marketplace), while the runtime has to live wherever your project's brepjs resolves. You install both once and never think about it again.

Install the skill (Claude Code plugin)

The skill ships through the brepjs plugin marketplace — which is just this git repo. Add the marketplace once, then install the plugin:

/plugin marketplace add andymai/brepjs
/plugin install brepjs-verify@brepjs

Claude Code now knows the workflow and will run the CLI for you. There is no npm step for the skill — plugins aren't loaded from node_modules.

Install the runtime (npm)

Add the CLI to the project where you want parts verified:

npm i -D brepjs-verify

That's the whole install. brepjs-verify bundles its own brepjs + occt-wasm, so it runs in an empty directory with nothing else set up. In an existing brepjs project it automatically prefers your locally installed brepjs and kernel (via a Node module-resolution hook), so your verified parts bind to the exact version you ship — no drift between what you verify and what you build.

It runs occt-wasm as its only kernel, rather than the auto-detect fallback chain, so verification results are reproducible on one known engine.

Prefer not to install anything? Run it straight from npm:

npx -y brepjs-verify part.brep.ts

The .brep.ts contract

A model is a TypeScript module whose default export is a zero-argument function returning a shape (or a Result<shape>):

// bracket.brep.ts
import { box } from 'brepjs';

export default () => box(40, 20, 10, { centered: true });

Declare intent with an expected block and the CLI asserts it — so you prove the part is the right part, not just a valid one:

// bracket.brep.ts
import { box } from 'brepjs';

export const expected = { volume: 8000, tolerancePct: 1 };
export default () => box(40, 20, 10, { centered: true });

Author parts in an ESM context (the CLI's default). A CommonJS project needs "type": "module" in package.json, or name the file .mts.

Quickstart

# scaffold a parameterized part
npx -y brepjs-verify init bracket

# author bracket/bracket.brep.ts, then verify (type-check + geometry) and write the report
npx -y brepjs-verify verify bracket/bracket.brep.ts --check --json report.json

# review it visually, then export the validated STEP
npx -y brepjs-verify verify bracket/bracket.brep.ts --snapshot shots/
npx -y brepjs-verify verify bracket/bracket.brep.ts --step bracket.step

The command exits non-zero whenever the report is not ok, so it drops straight into CI or an agent loop.

Why I built this

I write CAD as code, but I can't trust an agent to write it blind — and neither can the agent. Geometry that reads fine on the page is constantly wrong in ways only the kernel can see: a boolean that didn't overlap, a fillet that collapsed a face, a part that's valid but 3 mm too tall. So I gave the agent the one thing it was missing: a way to run the part, measure it, look at it, and find out the truth before claiming it's done. That loop is the whole idea — the rest is plumbing.

Next steps

• The Verify Loop — the author → verify → repair workflow and how to read the report
• CLI Reference — every subcommand, flag, and exit code
• Examples — the example gallery the skill learns from
• Eval & Scorecard — how the skill measures itself and stays honest