On the last day of October a repository appeared on GitHub: storytold/craft-anything. Ten days later it had around 25,000 stars, more than 5,000 forks, and version numbers from 0.1.0 to 0.6.0.
Its self-description is one sentence, and a strange one: see how any website, app or game works, then have AI build a version of your own.
The first reaction from most people is that this is another wrapper around a code generator. The README says otherwise. It splits looking and building into two phases, with an evidence ledger between them.
Four steps, each with a checkpoint
The loop is Study, Build, Check, Ready. Those are the four words on screen.
Study is where most of the engineering lives. The engine ships several hundred specialised probes grouped into seventeen toolsets, covering websites, native apps on Windows, macOS and Linux, Android and iOS, Electron and JavaScript apps, games, firmware, databases, binary formats and packet captures. For a website it measures layout at three screen widths, samples colours, identifies fonts, records spacing and captures menu states. For a program it watches runtime behaviour and file structure. Every finding is written as an evidence entry with a confidence level, a coverage statement and known limits; large result sets are summarised first and paged on demand.
Build hands the evidence to a coding agent. The design detail people miss: it does not ship a model. You connect an agent you already have signed in. Claude Code is first-class; OpenCode, Copilot CLI, Grok Build and anything that speaks ACP can be wired in. You choose the stack, and Rust, Unity, Unreal and Godot are what the checker recognises.
Check is what separates it from the rest. Instead of asking you to judge by eye, it puts the new project next to the original: websites compared at three widths, games running their own self-tests, apps compared by their runs. Output is a similarity score plus a list of what still differs. Ask in plain language for a fix and it runs another round.
Ready opens, plays or previews the result.
Clean room is the default, and the most debated part
Clean room means the code, files and page source of the original are never read. Tools that try are refused, with instructions for switching to Deep mode. What it may use is what an ordinary visitor can see: screenshots, visible text, measured layout, colours, fonts, spacing, and a window screenshot you supply yourself.
Deep mode unlocks everything, including decompilers, disassemblers and Ghidra integration, and is meant for software you own or are authorised to examine. On the command line, --clean-room or --deep switches for a single run and is not saved to settings.
The legal intent is obvious: clean room reverse engineering has a track record in copyright law because it is built to show independent creation rather than copying. The project also states its boundaries: it never bypasses DRM, logins or anti-cheat, and the agent is told not to copy logos, photos or text unless you confirm rights.
The timing was not accidental. Two days earlier, Anthropic restricted Claude Code from recreating existing games. craft-anything draws the opposite line: open source, outside provider control, with a compliance posture baked into the defaults.
The sixteenth tool from a tool shop
It comes from the ArtCraft team and belongs to the Craft App Suite, a family rewritten in Rust: PhotoCraft for images with real PSD support, FilmCraft for video, VectorCraft, PdfCraft, WordCraft, GridCraft, DeckCraft, SoundCraft, CADCraft. Those tools help you make things.
craft-anything is different. It does not make; it studies and rebuilds. The team's stated mission is to replace legacy SaaS one product at a time, and this repository is the sharpest, most legally exposed piece of that plan.
How the attention arrived
It did not trend on Hacker News: a November 5 submission scored 6 points. GitHub trending and non-English tech media did the work. A Korean tracker logged 400 stars on day one, a Japanese writer on note.com described the team as out to destroy legacy SaaS, and Chinese blogs published a dozen hands-on posts within two days.
Six releases between October 31 and November 10 is one every two days. That is both a sign of activity and a warning that things move. The README status line is plain: early alpha.
Should you use it now
If the goal is to rebuild a small site or utility you are allowed to study, it is usable today, provided you already run Claude Code. If the goal is a production-grade clone of a commercial product, not yet, especially outside the Claude path and across many convergence rounds, where the issue tracker documents real failures.
The next piece covers those failures and the limits the project admits to.
craft-anything Hit 25,000 Stars in Ten Days by Turning Software Research Into a Four-Step Pipeline
← Previous: Grok Bot Gets the Keys to Your Mac: Messages, Mail, Notes and Four More
Next: →The Other Side of 25,000 Stars: Three Kinds of Failure in craft-anything's Issue Tracker →