Point it at a website, game, or desktop app. It rebuilds the whole thing in Rust. Then it compares its output pixel by pixel against the original, grades itself, and keeps looping until the copy is faithful.
Free, open source, 35,000 GitHub stars in two weeks.
That is Craft Anything. Setup, real use cases for operators, and the limits to know before you rely on it are all below.
What Is Craft Anything?
Craft Anything is a Claude Code plugin built around a four-step loop: Study, Build, Check, Ready.
Study. The tool inspects the target with screenshots, layout measurements, colours, fonts, spacing, and opened menus. The default mode (Clean room) looks only at what a visitor can observe — no reading of internal file contents. Deep mode, which you opt into explicitly, also reads JavaScript source, disassembly, decompiled code, and call graphs. Use Deep mode on software you own or have permission to inspect.
Build. The tool passes the study data to Claude Code and builds a new project in the language you specify. Rust is the default. The tool supports websites, native desktop apps, .NET, Java, Android, Electron apps, games (Unity, Unreal, Godot), firmware, and more. 320 tools handle the inspection side. The tool includes its own disassembler, decompiler, and decoders without requiring external programs.
Check. After each build round, the tool runs a comparison against the original. For websites: screenshots at three screen widths, likeness scores per page, a list of exactly what does not match yet. For games: the game's own self-tests run against the original. The check results are stored in the project's checks/ subfolder and shown live in the app.
Ready. When the likeness score is high enough, the project is done. Or you type a change request and it refines.
The loop runs for one initial build plus up to six refinement rounds in the app (two in the CLI) by default. The --rounds flag changes that.
| Detail | Value |
|---|---|
| License | MIT / Apache-2.0 |
| Price | Free |
| Requires | Claude Code, signed in |
| Output default | Rust |
| GitHub stars | 39,000+ |
| Analysis modes | Clean room (default) / Deep |
Setup: 2 Options
Option 1: Pre-built installer (fastest)
Download the installer for your platform from the GitHub releases page. Available for macOS (signed, notarized, universal), Windows (x64, arm64, x86), Linux (AppImage, Flatpak, .deb, .rpm, tarball), and FreeBSD.
One note for macOS: download the CLI via curl rather than a browser. A browser download can trigger a Gatekeeper block ("Apple could not verify").
After installing, run:
Copy this.
craft-anything-cli doctor
This checks your install, confirms Claude Code is connected, and flags any missing pieces.
Option 2: Build from source
Requires Rust 1.97 or newer.
Copy this.
git clone https://github.com/storytold/craft-anything cd craft-anything cargo run --release -p craft-anything # desktop app # or cargo install --path apps/craft-anything-cli # CLI only
Linux desktop also needs build-essential libxkbcommon-dev libwayland-dev libx11-dev librandr-dev libxi-dev libgl1-mesa-dev libgtk-3-dev.
Connect to Claude Code:
Copy this.
craft-anything-cli setup
This finds the claude binary, connects the MCP server, and links the tool. Claude Code must already be installed and signed in first (claude auth login).
Done. Run your first rebuild:
Copy this.
craft-anything-cli rebuild https://example.com "Build a similar landing page"
What to Use It For
Understanding what a competitor built. Point Craft Anything at a competitor's web tool, pricing page, or marketing site. Clean room mode maps the layout, components, interactions, and visual style from what a visitor can observe. The check loop tells you how closely the study matches the actual implementation.
This is the fastest way to understand a competitor's frontend without guessing. You get a likeness score, a gap list, and a rebuilt version you can inspect and modify.
Rebuilding a UI for a client. A client wants something like a tool they already use. Instead of describing it from memory, point Craft Anything at the reference. It studies the real thing and builds from real evidence. The check loop closes the gap before you review.
Learning how an interface was actually built. The Advanced workbench shows function lists, pseudo-C, disassembly, call graphs, and an evidence ledger. For learning purposes, this is a faster path to understanding than reading source code.
Migrating a project to Rust. If you have an existing project you want to rebuild in a new language or framework, Craft Anything can study the existing version and start the build from real measurements rather than memory.
Game prototyping. Unity, Unreal, and Godot are explicitly supported. The check loop runs game self-tests against the original behaviour. Useful for rebuilding reference implementations or understanding how a mechanic works before building your own version.
What to Know Before You Trust It
Five limits worth knowing before you use this on real client work:
Behaviour checks for apps and games are not fully verified yet. The website pixel comparison is the most solid part. App and game behaviour checking is built but has not been fully tested against real programs on a live install.
Clean room is a tool rule, not an OS sandbox. In clean room mode, the tool will not read file internals. But a builder the user has allowed can still read and write as the user. Know what Claude Code has access to in your environment before you run a rebuild on a machine with sensitive files.
Installing dependencies runs code on your machine. Craft Anything asks for permission the first time it encounters a command type like
npmorcargo. Saying "always allow" for a type means every later command of that type runs without asking. Be deliberate about what you allow.The tool does not copy logos, photos, or text without confirmation. The agent is instructed to skip proprietary assets. For learning and analysis, that is fine. For client deliverables, check what you own or have rights to use before shipping anything.
Only Claude Code has been verified on a live install today. The tool lists support for Codex, Gemini CLI, Grok Build, Copilot CLI, and others. Most of those have not been run against a real install. Claude Code is the safe choice right now.
What It Actually Checks After Each Build Round
The check step deserves more detail because it is the part that makes this tool different.
For a website rebuild, the check takes screenshots at three different screen widths. It measures layout, colours, font sizes, spacing, and text content. Each page gets a likeness percentage. The project page lists exactly what does not match — a specific element that is offset, a colour that is off, a section that is missing. The agent uses that list to decide what to fix in the next round.
For a game rebuild, the check runs the game's own self-test suite against the original behaviour. For an app, the check compares the app's runs.
This is different from "it looks similar enough" — it is a scored comparison with a specific gap list. That gap list is what drives the refinement loop. By default it runs six rounds. At six rounds the tool stops and shows you what it has. You can then request more rounds or type a specific change.
The practical value: you do not have to eyeball the comparison yourself. The tool tells you the score and tells you what is left.
Pick one website you want to understand better. Run Craft Anything on it in clean room mode. The Study and Check steps show you more about how the site is built in fifteen minutes than an hour of DevTools.
The GitHub link is below. Read the responsible-use section in the README before your first run.