What Craft Anything Is: Study Anything, Then Build Your Own
Point Craft Anything at a website, an app or a game, say what you want, and it does the rest. That is the pitch at the top of the README, from the team behind ArtCraft: "See how any website, app or game works, then have AI build your own version." It is an open-source engine written in pure Rust, and it works in four steps. First it studies the original with real inspection tools. Then a coding agent builds a new project in the language, framework or engine you asked for. Then it checks the build against the original and lists what does not match yet. Then it refines the build until it is close, and you open, play or preview the result. The README is blunt about the state of the project: it is early alpha, with public releases and installers for Windows, macOS, Linux and FreeBSD, and it already passed 22,000 stars.
How It Works: Study, Build, Check, Ready
Every project follows the same four steps, and the app shows them on a project page in plain words. Study inspects the original: its pages at every width, its code, its files, how it runs. Build hands the job to a coding agent that builds a new project, in the stack your request names or a sensible default. Check compares the build with the original and lists what does not match yet. Ready is where you open, play or preview the result, or ask for changes and it refines it. The checks are specific to what you are rebuilding. A website is compared at three screen widths, a game runs its own self-test, and an app is checked by its runs. When two versions disagree, a Text Diff comparison of what each one produced is the fastest way to see exactly where they diverged. And because the checks run at three screen widths, an Aspect Ratio Calculator helps when you are sizing a rebuild for a phone screen versus a desktop one.
Clean Room by Default, Deep Mode When You Ask For It
The default is clean room, and the README explains exactly what that means. Nothing inside the original is read: not its code, not its files, not its page source. Craft Anything works from what you can see, the website as a visitor sees it, with screenshots, visible text, measured layout, colours, fonts and spacing, menus opened, and a picture of the app's window when you allow it. Tools that would look inside are refused, and they say how to turn on Deep mode. Deep mode is the opposite: every tool is used, for your private use. A program's code and files are read with the built-in engines, and Ghidra when it is installed, a website's JavaScript is read and its start page watched, and the builder may use all of them. Clean room is a rule about what Craft Anything offers and refuses. It is not an operating-system sandbox: it does not isolate your computer, and a coding agent you run still acts as you.
Built for Agents: 320 Tools, a CLI and an MCP Server
Craft Anything is built to be driven by agents as much as by people. The engine has 320 tools, and every one of them is also a CLI subcommand and an MCP tool, so any agent can drive the same engine the app uses. No AI is needed to inspect. From the command line, craft-anything-cli rebuild https://example.com "Recreate this website" runs a whole rebuild, and craft-anything-cli mcp serves the tools to any MCP client. Results come back as evidence records with their confidence, coverage and limitations, and big ones arrive as a summary first. Building needs a coding agent on your computer. Claude Code is supported first: Craft Anything runs your own, unmodified claude program, which uses your own sign-in, and Craft Anything never sees or handles your account. GitHub Copilot works through its installed CLI, Grok Build through its installed program, and other agents that speak ACP can be switched to it per backend. There is also a Simple mode and an Advanced mode. Simple asks "What do you want to make?" and shows the four steps in plain words. Advanced opens the full workbench: every tool, functions, strings, pseudo-C, disassembly, hex, call graphs, evidence, unknowns, jobs, logs and the agent console.
Getting It Running: Rust, One Command, and a Doctor Check
Getting started needs Rust 1.97 or newer, because a dependency requires it. On Debian or Ubuntu the desktop app also needs a handful of system libraries for Wayland, X11 and GTK. Then it is a clone and a run: git clone https://github.com/storytold/craft-anything && cd craft-anything, then cargo run --release -p craft-anything for the desktop app, or cargo install --path apps/craft-anything-cli for the CLI. The command craft-anything-cli doctor checks the install, the coding agents and the engines, and says what is missing. New projects go in Documents/Craft Anything, and each project folder holds craft.json with the request and status, capture with what was recorded from the target, app with the generated project, checks with each comparison against the original, and inbox with open questions. A project can also start in a folder you pick, and if that folder already has files you say what they are: the app itself, which the builder edits in place, or material to start from.
What It Does Not Do: The Rules and the Limits
The README puts its rules up front, and they are worth reading before you point it at anything. Use Craft Anything on software you own or are allowed to examine. It never bypasses DRM, logins or anti-cheat. Rebuilds recreate layout, style and behaviour from evidence, and the agent is told not to copy logos, photos or text unless you confirm you own them or have permission. Installing dependencies runs code on your computer, and the app asks before a command of a kind it has not seen in the project, so only allow a package manager freely for a project whose dependencies you trust. The honest limits are in the status section. What works today is the desktop app, the CLI and the MCP server, the Study to Build to Check to Refine workflow for websites, apps and games with Claude Code as the builder, website checks at three screen widths, and game rebuilds in Unity, Unreal or Godot. Partly there are building with API keys and other coding agents, which are built but not yet run live, and behaviour checks for apps and games. Generated pictures and 3D models are opt-in and need your own fal.ai key. It is early alpha, and the README says so plainly, but the workflow it promises already runs end to end.



