Point Craft Anything at a website, desktop or phone app, or game, tell it what you want, and let it go to work. Built 100% in Rust, this open-source engine inspects the original target using 320 real analysis tools, drives a coding agent on your machine to build a new project in your chosen stack, and checks the result against the original to refine what does not match yet.
How It Rebuilds Without Guessing
When you ask an AI to recreate a website or app from a prompt alone, it guesses. Craft Anything replaces guessing with a four-step loop: Study, Build, Check, and Ready. During the Study phase, inspection tools examine the target—its pages across screen widths, its files, its code, and how it runs—collecting structured evidence records behind every finding.
Next, a coding agent builds the new project in the language, framework, or engine you requested. Then comes the Check step: every build is compared directly against the original (such as testing web pages at three screen widths, running a game's self-test, or checking an app's execution) to list what still differs so the coding agent can refine it.
Clean Room vs. Deep Mode
Craft Anything separates inspection into two distinct modes depending on what you need:
Clean room (default): Nothing inside the original target is read—no source code, internal files, or page source. The engine works strictly from what a user or visitor can see: screenshots, visible text, measured layout, colors, fonts, spacing, opened menus, or an allowed picture of an app window. Any tool that would look inside is refused.
Deep mode (for private use): All 320 tools are unlocked. Built-in decoders, disassemblers, and decompilers (plus Ghidra if installed) read program code and files, inspect website JavaScript, and feed that evidence to the builder.
One important caveat: Clean room is a rule governing what tools Craft Anything offers and refuses, not an operating-system sandbox. It does not isolate your computer, and any coding agent you run still operates with your user permissions.
What Is Under the Hood?
At its core, Craft Anything is a pure-Rust reverse-engineering engine with 300 engine tools (and 20 more in the AI and setup layers) covering native binaries, .NET, Java, Android, Apple, Electron/JavaScript, websites, firmware, packet captures, databases, and game archives. Its decoders, disassembler, and decompiler run without external dependencies, and you do not even need AI to use them. You can drive the engine through the Simple four-step view, the Advanced reverse-engineering workbench (built with egui), the CLI (craft-anything-cli), or as an MCP server.
Feature | Specification in Craft Anything |
|---|---|
Language & UI | 100% Rust (egui desktop app, plus WebAssembly support for browser components) |
Platforms | macOS, Windows, Linux, FreeBSD, and Web |
Tool Count | 320 tools (300 in the core engine + 20 in AI and setup layers) |
Interfaces | Simple mode, Advanced workbench, CLI subcommands, and MCP server |
License & Status | MIT OR Apache-2.0 — Status: early alpha |
Build Requirement | Rust 1.97 or newer |
Which Coding Agents Work With It?
To build projects, you need a coding agent signed in on your own machine. Claude Code is supported first, running your unmodified local binary under your own account without Craft Anything touching your credentials. Other agents that speak ACP—including GitHub Copilot CLI, Gemini CLI, Qwen Code, opencode, goose, Hermes, and Maki—can be connected per backend, though only Maki has been tested live so far, and the Codex adapter has not yet been verified on a real install. For questions and analysis, it supports Anthropic, OpenAI, OpenRouter, and local OpenAI-compatible servers.
Before You Get Excited
The project is explicitly in early alpha. Building from source requires Rust 1.97 or newer, and automated rebuilding relies heavily on external coding agents and their token limits. Several agent adapters are still unverified in live environments.
My Take
The most compelling part of Craft Anything is not just cloning a webpage, but wiring a 300-tool Rust reverse-engineering engine directly into an MCP server and an automated verification loop. Having self-contained disassemblers, firmware parsers, and packet analyzers accessible as structured CLI and MCP commands is genuinely useful for engineering workflows even if you never touch the "rebuild" button. If I were to test it first, I would skip the simple game demos and point Deep mode at a real firmware image or native binary via MCP to see how accurate its built-in decompiler is compared to Ghidra. My main doubt is how well coding agents handle complex software rebuilds without getting lost in state, though the fact that large evidence records return as summaries first and page through core_evidence_browse shows thoughtful engineering around context limits.
