Description
You find a site, a small tool or a game you like and want your own version, but you can't tell how it's put together, let alone write it from scratch. Craft Anything hands that to AI in four steps: study the original, have a coding agent build a new project, compare the result against the original, then refine what doesn't match. You paste an address, say what you want, such as "recreate this website" or "a snake game in Rust", and press Start.
It doesn't write code from impressions. Websites are screenshotted at three screen widths and scored for how alike they are; mismatches come back as a checklist; and before the agent runs a command it asks you in an approval card. The engine is pure Rust, shared by the desktop app, the CLI and the MCP server.
Study the target: websites, desktop programs, phone apps, games and files can all be analysed, with the evidence behind each answer rather than a guess.
Clean room and Deep modes: the default clean-room mode never reads the original's code or files, only what an ordinary visitor sees: pages, text, layout and colours. Deep mode reads code, for software you're allowed to examine, and can use Ghidra.
Build and verify: new projects come out in the language or framework you ask for, games in Unity, Unreal or Godot, and each round is checked against the original until it's close.
Simple and Advanced views: Simple walks the four steps; Advanced is a full workbench with function lists, pseudo-C, disassembly, hex, call graphs and an evidence ledger.
Built for agents: 320 tools, each a CLI subcommand and an MCP tool, so another agent can drive it without the UI.
Bring your own coding agent: building uses a signed-in Claude Code on your machine, with GitHub Copilot CLI and others pluggable; Anthropic, OpenAI, OpenRouter or local compatible servers can answer questions.
Platforms and licence: installers for Windows, macOS, Linux and FreeBSD, Apache-2.0, still early alpha; the authors ask that you use it only on software you own or may examine, and it doesn't bypass DRM, logins or anti-cheat.
It doesn't write code from impressions. Websites are screenshotted at three screen widths and scored for how alike they are; mismatches come back as a checklist; and before the agent runs a command it asks you in an approval card. The engine is pure Rust, shared by the desktop app, the CLI and the MCP server.
Features
Study the target: websites, desktop programs, phone apps, games and files can all be analysed, with the evidence behind each answer rather than a guess.
Clean room and Deep modes: the default clean-room mode never reads the original's code or files, only what an ordinary visitor sees: pages, text, layout and colours. Deep mode reads code, for software you're allowed to examine, and can use Ghidra.
Build and verify: new projects come out in the language or framework you ask for, games in Unity, Unreal or Godot, and each round is checked against the original until it's close.
Simple and Advanced views: Simple walks the four steps; Advanced is a full workbench with function lists, pseudo-C, disassembly, hex, call graphs and an evidence ledger.
Built for agents: 320 tools, each a CLI subcommand and an MCP tool, so another agent can drive it without the UI.
Bring your own coding agent: building uses a signed-in Claude Code on your machine, with GitHub Copilot CLI and others pluggable; Anthropic, OpenAI, OpenRouter or local compatible servers can answer questions.
Platforms and licence: installers for Windows, macOS, Linux and FreeBSD, Apache-2.0, still early alpha; the authors ask that you use it only on software you own or may examine, and it doesn't bypass DRM, logins or anti-cheat.
