中文版

craft-anything（GitHub：https://github.com/storytold/craft-anything ，2026-10-31 创建，今晨抓取时已 11,788★）是 ArtCraft 出品的纯 Rust 开源引擎，口号是“理解任何软件如何工作，再做出你自己的版本”：指着一个网站、应用或游戏，说一句想要什么，它先用真实的检查工具研究原作，再让编码 Agent 建一个新项目，最后把结果和原作逐项比对、打磨到接近。

它的流程是四步：Study（检查原作的页面、代码、文件与运行方式）→ Build（编码 Agent 按你指定的语言、框架或引擎建新项目）→ Check（和原作比对，列出还不匹配的地方）→ Ready（打开、试玩或预览结果，还能继续提修改）。默认是 Clean room 模式：不读原作的代码、文件或页面源码，只看访客可见的东西（截图、可见文字、实测布局、颜色、字体与间距）；Deep 模式才会动用全部检查工具，包括内置解码、反汇编与反编译，装了 Ghidra 时也会用上。整个引擎有 320 个工具，每个工具既是 CLI 命令也是 MCP 工具，所以任何 Agent 都能直接驱动它；构建默认走你本机已登录的 Claude Code，也支持 Codex、Copilot CLI 等走 ACP 的 Agent。桌面应用、CLI 与 MCP 服务器都有，Windows、macOS、Linux、FreeBSD 与 Web 版齐全，但作者明确标注仍是早期 alpha。

为什么值得了解

• 对正在学逆向、学 Agent 编排的开发者，这是一个“看得见过程”的学习样本：研究、构建、比对、修正四步全是显式步骤，而不是一句提示词交出去就等结果
• “先比对再交货”的思路值得借鉴：网站在三种屏幕宽度下比对、游戏用自测、应用看实际运行，检查项不合格就继续改，而不是凭感觉说完成了
• 但要按作者自己的提醒来用：只用在你拥有或被允许检查的软件上，它不绕过 DRM、登录或反作弊；重建的是布局、风格与行为，Logo、照片与原文除非你确认拥有或获得许可，否则 Agent 被要求不照搬

English Version

craft-anything (GitHub: https://github.com/storytold/craft-anything , created 2026-10-31, 11,788 stars at this morning’s fetch) is ArtCraft’s pure-Rust open-source engine whose motto is “understand how any software works, and craft your own”: point it at a website, app or game, say what you want, and it first studies the original with real inspection tools, then has a coding agent build a new project, and finally compares the result against the original item by item and refines it until it is close.

Its workflow has four steps: Study (inspect the original’s pages, code, files and runtime behaviour), Build (a coding agent builds a new project in the language, framework or engine you name), Check (compare with the original and list what does not match yet), and Ready (open, play or preview the result, or ask for changes and it refines further). Clean room mode is the default: it reads none of the original’s code, files or page source, working only from what a visitor can see (screenshots, visible text, measured layout, colours, fonts and spacing); Deep mode brings in the full inspection toolkit, including built-in decoders, a disassembler and a decompiler, plus Ghidra when installed. The engine has 320 tools, each one both a CLI command and an MCP tool, so any agent can drive it directly; building defaults to your locally signed-in Claude Code, with Codex, Copilot CLI and other ACP-speaking agents also supported. Desktop app, CLI and MCP server are all available for Windows, macOS, Linux, FreeBSD and the web, but the author clearly labels it early alpha.

Why it’s worth knowing

• For developers learning reverse engineering or agent orchestration, this is a sample where the process is visible: study, build, check and refine are all explicit steps, not a single prompt handed off and waited on
• The “compare before delivering” idea is worth borrowing: websites are compared at three screen widths, games use their own self-tests, apps are judged by actual runs, and anything that fails a check gets refined rather than declared done on feeling
• Use it with the author’s own warning attached: only on software you own or are allowed to examine — it does not bypass DRM, logins or anti-cheat; rebuilds recreate layout, style and behaviour, and logos, photos and text are not to be copied unless you confirm you own them or have permission