Craft Anything rebuilds a website with Claude Code and checks on its own how close it gets to the original, up to 90%

Craft Anything is a free, open-source program from ArtCraft that rebuilds an entire website with Claude Code and then checks on its own how closely the new version matches the original. It captures the site on desktop, tablet and phone, compares, sends the fixes to Claude Code and repeats until the similarity reaches 90%. It came out on October 31, 2026, and passed 10,000 stars on GitHub in 4 days.

It comes from the same team that had already rewritten Photoshop and the other Adobe apps from scratch: I covered here how ArtCraft rewrote Adobe and Office in clean-room.

What changes compared to rebuilding a site by hand with Claude Code?

Until yesterday, rebuilding a site with Claude Code meant taking screenshots by hand and explaining every detail. Claude Code worked partly on guesswork, and checking the result was on you.

Craft Anything automates the two most tedious parts: collecting what is visible on the original site and verifying how closely the copy matches it. Claude Code is still the one writing the code.

Can AI now clone any software?

No. A few weeks ago I explained that copying software through reverse engineering has at least three limits, and they still hold.

Craft Anything works differently: by default it doesn’t copy the site’s code. It looks at the site exactly as a visitor sees it and builds a new version from there.

How Craft Anything works, step by step

1. Pick what to rebuild. Open the app and choose a website, an app or a game.
2. Say what you want. Enter the address and a request, for example "recreate this website".
3. Choose the mode. Clean Room or Deep (see below).
4. Press Start. The program opens the site the way you would and captures it on desktop, tablet and phone.
5. It studies the site. It records colors, fonts and spacing, opens the menus and captures up to 4 other pages of the site.
6. Claude Code builds. All that material goes to Claude Code, which develops your version in a new folder.
7. It checks and fixes. Craft Anything captures your version too and compares it with the original.

According to the project documentation, the comparison runs at three screen widths: 1440, 768 and 390 pixels, meaning desktop, tablet and phone.

What’s the difference between Clean Room and Deep mode?

Clean Room mode is the default. Craft Anything only looks at what any visitor would see: screenshots, visible text, layout, colors, fonts, spacing and opened menus. The code and the page source are not read.

Deep mode also reads the code, for example the site’s JavaScript. The authors say to use it for private use only.

One clarification from the README: Clean Room is a rule about what Craft Anything offers and refuses to do. It doesn’t isolate your computer, and Claude Code still acts with the permissions you gave it.

How does Craft Anything know when the site is ready?

This is the interesting part. After each round Craft Anything tells you how alike the two versions are and what is still missing: a heading, a color, a wrong margin.

Then it sends that list back to Claude Code, which fixes it. The loop keeps going until it reaches 90% similarity. The documentation caps it at 6 refinement rounds in the app: it stops at whichever limit comes first.

That way the check no longer depends on your eye. There’s a number that tells you how far off you are, and a precise list of what to fix.

What do you need to use Craft Anything?

• Claude Code installed on your computer. The Claude desktop app isn’t enough: the program flags this exactly when it only finds the desktop app.
• A Claude subscription with room to spare. A rebuild eats a good part of your subscription.
• The program. Installers for Windows, macOS, Linux and FreeBSD are on the Releases page of the GitHub repository.

Claude Code is the first supported agent. The README also mentions Codex and GitHub Copilot, but as integrations that haven’t been fully verified yet.

What are Craft Anything’s limits today?

Craft Anything is still an alpha, so it’s rough. For now the comparison with the original only covers websites: for apps and games the behavior checks are still partial.

And there’s one limit that doesn’t depend on the software. Use it only on your own sites or on sites you have permission to study.

Who should try it now

The most natural use case is an old site to rebuild with modern technology, starting from what is already visible online. The loop that captures, compares and fixes is a good example of how to orchestrate an agent: you give it a measurable goal and it iterates until it gets there.

If you have a 20-year-old site to bring back to life, you now have a concrete way to start.