CADAM: Open-Source Text-to-CAD with Parametric Sliders and WebAssembly Rendering
Adam (YC W25) released CADAM, an open-source text-to-CAD web app that generates parametric 3D models from natural language or image references. It outputs OpenSCAD code with automatically extracted parameters that surface as interactive sliders for instant dimension tweaking. You can try it live at adam.new/cadam.
Key Features
- Two generation modes: Parametric mode writes/edits OpenSCAD via a
build_parametric_modeltool; mesh mode generates textured 3D meshes. - Deterministic slider updates: Adjusting a slider does a regex-based update on the SCAD source — no LLM call needed for simple parameter changes.
- Model-agnostic backend: Uses Vercel AI SDK to support Anthropic (Claude), Google (Gemini), Gemini 3.1 Pro (top performer in evals), and OpenAI/others via OpenRouter. Adaptive thinking auto-enables on newer models.
- Browser-based rendering: Compiles OpenSCAD to WebAssembly in a Web Worker, UI never blocks. Renders with Three.js via React Three Fiber.
- Library support: Includes BOSL, BOSL2, and MCAD. Built-in Geist font for text in models.
- Export formats: .STL, .SCAD, .OBJ, .GLB/GLTF, .FBX, .DXF.
Under the Hood
Built with React (TanStack Start) and Supabase for auth, database, and file storage. The repo includes benchmarks for complex assemblies — V8 engine, radial engine, turbofan jet engine — all generated from single prompts with multiple parametric controls.
Future Plans
- Add build123d and CadQuery for constraint-driven modeling beyond CSG primitives.
- Better spatial context: UI for face/edge selection and viewport image integration for LLMs.
Clone the repo from github.com/Adam-CAD/CADAM and run locally. Contributions welcome.
📖 Read the full source: HN AI Agents
👀 See Also
docvault: Generate Local API Docs to Reduce AI Hallucinations
docvault is a tool that generates markdown API references from source code to help Claude and other LLMs stop hallucinating function signatures. It works for Rust crates and Python packages, outputs a two-tier markdown file, and includes a Claude Code plugin for hands-free operation.
Anthropic's Multi-Agent Harness Design for Improving Claude's Code Quality
Anthropic's blog post details a harness design using multiple agents to address Claude's context anxiety and self-evaluation bias, with specific agent roles and scoring criteria for frontend and full-stack development.
The Bottleneck in Parallel AI Agents: Human Approval Queue Bottleself
A developer running parallel Claude Code agents describes the 'bottleself' — the point where parallelism stops adding output and starts creating a backlog of human approvals. Their solution: a planner that decomposes goals into subtasks, spawns agents, and only pings on unresolved decisions.
Claude Code Plugin for Reddit Business Research
A Claude Code plugin automates Reddit research for businesses by searching relevant posts, analyzing threads, and generating structured markdown reports with findings and source links. No API keys required—install via GitHub and run with a single command.