open-design

I'm building an open-source, local-first design tool that works as an alternative to Claude Design. It's a web-deployable platform that auto-detects coding agents on your system (Claude Code, Cursor, Gemini CLI, Copilot, etc.) and uses them as the design engine. The core idea is that you describe what you want to design—websites, mobile apps, desktop prototypes, slides, images, videos—and the AI agent generates it for you.

The architecture should be a monorepo using pnpm workspaces with TypeScript throughout. I need multiple packages that handle different concerns: one for the web interface, one for agent detection and CLI spawning, one for the design system management, one for skills (composable design operations), and one for export handling (HTML, PDF, PPTX, MP4). Think of skills as modular, reusable design operations—like "create button," "generate layout," "add animation"—that can be chained together.

The platform should support 70+ pre-built design systems (like Material Design, Tailwind, etc.) that users can apply to their generated designs. There's a sandboxed preview system so generated code runs safely before export. Users should be able to export their work in multiple formats.

For the agent integration, auto-detect what's available on the user's PATH and spawn it as a subprocess, passing design requests as structured prompts. If no local agent is found, fall back to an OpenAI-compatible BYOK proxy. The CLI should be the primary interface, but also support web-based workflows.

The project is heavily documented with i18n support (multiple languages), has Docker support, and follows a contribution-friendly structure. Use modern TypeScript practices, keep the codebase modular so skills and design systems are easy to add, and make sure the export pipeline is sandboxed and secure. The whole thing should feel fast and responsive, whether someone's using it via Claude Code or through a web interface.