v0.2.0

Use it from your coding agent. breakinggood mcp runs an MCP server over stdio with three tools: capture_site (what a page is made of: tokens, outline, DOM, API calls, screenshot paths), scan_repo (stack + dependency risk report) and clone_and_prepare (clone, scan, .env, install — returns the run command).

claude mcp add breakinggood -- npx -y breakinggood mcp

Also: a real demo recording in the README, generated projects now credit the source URL in their README, and the dependency scanner no longer flags a repo's own workspace packages.

v0.1.0

First release.

npx breakinggood https://github.com/someone/some-app   # clone, scan deps, set up .env, install, run
npx breakinggood https://some-website.com              # capture it, rebuild as Vite + React + Tailwind, run

Repo mode detects Node (npm/pnpm/yarn/bun), Python, Go, Rust, Ruby, PHP, Java, .NET, Docker Compose and static HTML. Before installing it reads the lockfile for install scripts, foreign registries and brand-new packages, looks for curl | bash and obfuscated blobs, and stops to ask on anything high-risk. .env.example becomes .env with random secrets filled in.

Site mode drives your installed Chrome/Edge headless, captures tiled screenshots, a cleaned DOM, computed design tokens and every XHR/fetch with its JSON, then has a model (Claude, OpenAI, Groq, OpenRouter, Gemini or local Ollama — whichever key you have) write a runnable project with a mock API that replays the captured responses.

Tested on Windows, macOS and Linux with Node 20 and 22. See the README for the before/after shots of Hacker News and tailwindcss.com.