The Signal

A solo developer shipped FluidCAD — a parametric CAD tool where you write JavaScript to build 3D models. Built over a year in free time. It runs on OpenCascade.js, a WASM binding to the same BRep kernel used in serious engineering software. You get fillets, chamfers, STEP import/export. The editor lives in your local files — use VS Code, Neovim, whatever you already use . The browser gives you a live-updating webview as you type. It auto -fuses intersecting geometry so you don't have to manually manage boolean operations. Feature-level transforms ( not just shape transforms) are the headline differentiator.

Builder's Take

Here's the leverage calculation: parametric CAD has always been locked behind GUI-first workflows. Fusion 360, S olidWorks — they're built around mouse interaction. That's a 10x friction multiplier for anyone who thinks in code .

FluidCAD flips the model. Your design is your source code. That means :

  • Version control is free. Git diff your CAD. Branch a design variation. Merge mechanical parts. No GUI tool gives you this .
  • Programmatic generation is trivial. Want 50 variations of a bracket with different hole spacings? Write a loop. This is the killer use case for AI- assisted hardware design.
  • LLM integration is obvious. If your CAD is just JavaScript, you can prompt GPT-4o or Claude to generate or modify models. The code is the interface.

The moat this creates: a solo dev building physical AI products (enclosures, mounts, fixtures, custom hardware) can now script their way through mechanical design without leaving their terminal . The moat this destroys: any SaaS that charges for "parametric scripting" as a premium tier.

The auto-fuse behavior is under rated. In traditional CAD, boolean union management is where begin ners break. Removing that cognitive load means faster iteration cycles. For a one-person team shipping hardware, fewer context switches = more shipped .

Tools & Stack

FluidCAD

  • URL: fluidcad.io
  • Pricing: Check current pricing — early-stage project, likely free or freemium at launch
  • Kernel: OpenCascade.js (WASM) — same BRep engine as FreeCAD and many industrial tools
  • File formats : STEP import/export (the universal standard for hardware handoff to manufacturers)
  • Editor: Your local editor + browser webview. No lock-in.

Comparable Tools

  • OpenSCAD — the OG code-based CAD. Uses its own language (not JS). Mature , but the syntax feels dated and the renderer is slower. No BRep — it's mesh-based, so no fillets without hacks.
  • CadQuery — Python-based, BRep via OpenCascade. Serious engineering use cases. Steeper setup curve, no live browser preview out of the box.
  • OpenJSCAD — JavaScript CAD, but mesh-based (CSG). No fillets, no STEP export.
  • Fusion 360 — Has a scripting API (Python/JS), but you 're scripting inside a GUI app. Not the same as code -first.

FluidCAD's edge: JavaScript + BRep + live preview + local file editing. That combination doesn't exist anywhere else right now.

Quick Start Pattern

Based on the documented workflow, a basic model looks roughly like this:

// Conceptual FluidCAD pattern  (check docs for exact API)
const box  = Box({ width: 50, height: 30, depth: 20 });
const hole = Cylinder ({ radius: 5, height: 30 });
  .translate([25, 15, 0]);
 
// Objects that intersect are auto-fused — no  manual boolean needed
export default [box, hole];

Check the lantern tutorial and patterns guide for real working examples.

Ship It This Week

Build an LLM-to-CAD Micro-Tool

Here's the concrete project: build a small web app where a user describes a simple part in plain English, your backend sends that to the OpenAI or Anthropic API asking it to produce valid FluidCAD JavaScript, and the output is displayed as runnable code the user can drop into FluidCAD.

Stack:

  • Next.js or plain Express for the API route
  • Claude 3.5 Sonnet or GPT-4o with a system prompt that includes FluidCAD's API reference as context
  • A textarea showing the generated JS + a copy button
  • Optionally embed the FluidCAD webview directly if they open -source the renderer

Why this works: FluidCAD's JS output is small, structured, and constrained — exactly the kind of code LLMs generate reliably. You're not asking the model to write a web app. You're asking it to call a small set of geometry primitives. Token count stays low. Accuracy stays high.

Monetization angle: Charge makers, 3D printing hobbyists, and hardware founders $9/month for 100 AI-generated parts. The niche is tiny and under served. Pieter Levels built a $40K/ month business on a niche this size.

Start now: Clone the FluidCAD docs, paste the API reference into a Claude Project or a GPT-4o system prompt, and test 10 prompts manually. If 7/10 produce valid-looking code, you have a product.