Craft Anything Tested: The Free REA Alternative Rebuilt Our Website From Screenshots Alone

On October 10, Claude stopped itself halfway through taking apart an app for us. Over the next two weeks a lot of people hit the same wall, and they started saying Claude had blocked REA (short for Reverse Engineer Anything), the most popular tool for this job. Then the team that rebuilt every Adobe app with Claude shipped their answer: Craft Anything. It's free and open source.

We tested it on three jobs, with REA as a control. The short answer: Craft Anything rebuilt a real app feature with a perfect 4,915 out of 4,915 score and our whole website from screenshots alone at 91% alike. But it is not a way around Claude's check, and it was not the cheapest option.

What Craft Anything does

You point Craft Anything at something: a website, an app on your computer, a phone app, or a file. You say what you want in plain words, like "rebuild this search box", and you press Start. Then it works in four steps:

1. Study. It looks at the original with real inspection tools.
2. Build. A coding agent writes your own version, in whatever language or framework you ask for.
3. Check. It compares your version with the original and lists what doesn't match.
4. Ready. You open it, or you ask for changes and it goes again.

Under the hood there are 320 tools, all written in Rust. Every one is also a command line tool and an MCP tool (a standard way for AI agents to call outside tools), so Claude Code, Codex or any other agent can use it without the app. When we installed it, its doctor check found our Claude Code on its own and said healthy. No API key needed.

Clean room vs Deep

The most interesting part is one switch.

• Clean room means it never looks inside the original. It only uses what any visitor can see: screenshots, layout, colors, fonts and spacing.
• Deep means it also reads the app's code and files, and the README says that's for your private use.

And here's the part the viral posts left out. Craft Anything doesn't replace Claude. It runs your own copy of Claude Code, signed in with your own account. Which raises the obvious question: if it's still Claude underneath, why would it work when REA didn't?

Was REA really blocked?

Anthropic never named REA. What happened is that their newest model, Opus 5.5, shipped with stronger safety checks for cyber work. The message it showed us in October says those checks "can sometimes flag non-cybersecurity work", and it offers to switch to the older Opus 4.8 and keep going.

Separately, Claude Code reportedly now steers away from recreating specific existing games. That's a different rule, and we kept games out of all our tests.

Test 1: the job Claude stopped

Same app as October: Obsidian, the note-taking app. Same question: list every message its window can send to the main program, and what each one does. Same model, Opus 5.5. Only the tool is different.

It went straight through. No stop. Craft Anything found all 42 channels, grouped them, and showed where each one lives in the code, in 3 minutes 9 seconds.

One detail we really liked: for every channel it marked whether the answer was supported by the actual code, or only inferred from the name. Sixteen of them were guesses, and it told us so, and listed what it still didn't know. Most AI tools never tell you which parts they made up.

But we wanted to be fair. So we ran plain REA again, the same day, with the exact prompt from October. It went through too. Same 42 channels. No stop.

So the honest answer is: Craft Anything didn't get around anything. Claude's check has changed since October. The real question is which tool does the job better.

Round 1: speed

On that same job, Craft Anything took 3 minutes 9 seconds. REA took 12 and a half minutes, because its full analysis of the app ran into its own five-minute limit and had to start again on a smaller part.

Round 2: rebuilding a real feature

Obsidian's quick switcher is the search box that finds any note as you type. We asked Craft Anything, in Deep mode, to work out how it ranks results and rebuild it as a web page. Then we checked the rebuild against Obsidian's own search code with 4,915 searches.

Every single one matched, in the exact same order. It also wrote a document explaining the scoring rules, and every number in the page comes from that document. It even explained the quirks, like why typing just the two letters D and N puts a note called Reading List at the top. It found the exact rule in Obsidian's code and showed us where it lives.

But it wasn't the fastest or the cheapest. It took 6 minutes 1 second and about $3.59 at API prices. When we ran REA on the same job in October, it took 4 minutes 11 seconds and about $1.48. And a lighter setup, LeanREA, did it for $0.52.

Round 3: our website from screenshots

This is the one REA isn't built for. Clean room. We gave it our own website, hyperautomationlabs.co, and told it not to look at any code. Screenshots only. Then we let it run.

It checked itself after every pass. 84% alike. 86. 88. Then it noticed the blog page was dragging the score down, redrew 22 thumbnails by hand, and passed its own target at 91%.

Side by side: same headline, same newsletter box, same review card, and the page height is off by two pixels. Phone and tablet came in at 85%. And it didn't stop at the home page. It rebuilt five pages, including the blog and the compare pages, with the same menu linking them. On phones, the menu turns into a button.

Costs are Claude Code's API-price estimates. On a Claude Max subscription it's usage, not a bill.

The catch

The pictures in the rebuilt site are hand-drawn stand-ins, not our real thumbnails. And the job took 36 minutes, which Claude Code priced at about $24 at API rates. On a Claude Max subscription that's part of your usage, not a bill, but it's a lot of usage.

Why so much? After every pass it takes screenshots at three screen widths, compares every page, and keeps refining until it hits its 90% target. That's great for accuracy, and heavy on usage.

One more thing you should know. We ran it in unattended mode, and it decided on its own that we owned the site because our email matched the domain. That was right here, but we'd rather it asked.

Who should use it, and setup

If you see a website or app feature you love and want your own version in your own stack, Craft Anything is the most complete tool we've tried. Start in Clean room, because it never touches the code of the original.

If you only need answers about how an app works, REA or a lighter setup is cheaper. Use Deep mode only on apps you own or have the right to study, and keep the result private.

It's labeled early alpha, so expect rough edges, and watch your usage on big jobs. Setup takes five minutes:

1. Download it from the official GitHub releases page.
2. Sign in to Claude Code in your terminal.
3. Pick a target and press Start.

Get the Setup Card

We put the setup steps, our exact prompts from all three tests, and which mode to use when in one free card. Comment CLEANROOM on the video, or grab it below. And if you want to see where all of this started, watch how the same team used Claude to rebuild every Adobe app.

How we tested: Craft Anything v0.3.0 running our own Claude Code with Opus 5.5, November 5, 2026. REA control run on the same day with the October prompt; REA feature-rebuild numbers are from our October run. Costs are Claude Code's API-price estimates, not bills. One small test, not a study. Independent, not sponsored by or affiliated with ArtCraft or Anthropic.