Craft Anything: Free Reverse Engineering That Doesn't Ask Claude's Permission

TL;DR: Craft Anything is a free, open-source engine from the ArtCraft team that studies a website, desktop app, phone app or game with its own inspection tools, then has a coding agent build your own version and checks the result against the original. The inspection half needs no AI at all: 314 tools ship as a pure-Rust engine, a CLI and an MCP server. It is dual-licensed MIT or Apache-2.0, sits at v0.3.0, passed 6,100 GitHub stars within days of release, and is openly labelled early alpha.

On 30 October 2026, Anthropic added two sentences to Claude Code's system prompt. One told it to build original games, genre games, and games from the user's own designs and assets. The other told it not to build a recreation of a specific existing game, "even if it is renamed or called original." Addy Osmani, a member of Anthropic's technical staff, confirmed the change on X that day and said using Claude to infringe IP is against the company's usage policy.

The stated scope was narrow. What people reported was wider. Tom's Hardware says users, including one of its own editors, hit refusals on ordinary reverse-engineering work too, and that switching back to an older Claude model finished the job. The Hacker News thread under the story ran to 121 comments and the mood was not warm.

Craft Anything landed three days later, from the team behind the ArtCraft apps. Its pitch is blunt: point it at a target, see how it works, then build your own. The interesting part is not that it works around a refusal. It is that the reverse engineering is done by a Rust engine on your own machine, by 314 tools that never ask a model for permission.

What Craft Anything Actually Does

The simple mode is one screen. Pick a website, an app or a game, give it an address, describe what you want, press Start. From there it runs four steps and narrates them in plain words: Study, Build, Check, Ready.

1. Study. Inspection tools look at the target and record what they find, with the evidence attached to every answer.
2. Build. A coding agent writes a new project in the language, framework or engine your request names.
3. Check. The build is compared against the original. Websites are compared at three screen widths, a game runs its own self-test, an app's runs are compared.
4. Ready. You open, play or preview the result, then ask for changes and it refines.

Approval cards interrupt when the agent wants to do something outside the rules you set. Anything that falls outside the sandbox waits for your click. There is also a "just look inside" path that opens a target for questions without building anything, which is the mode that matters for anyone who only wants to know how a thing works.

An Advanced workbench exposes the machinery: function lists, strings, pseudo-C, disassembly, hex, call graphs, an evidence ledger and the raw agent console. The simple app and the workbench drive the same engine, so nothing is a demo mode.

The AI Builds, the Engine Inspects

This split is the whole product argument, and it is worth being precise about it.

Each of the 314 tools is both a CLI subcommand and an MCP tool. The README says it outright: no AI is needed to inspect. The disassembler, decompiler and decoders are the project's own Rust code, covering native binaries, .NET, Java, Android, Apple, Electron and JavaScript, websites, firmware, packet captures, databases and game files. Ghidra gets used in Deep mode when you have it installed, and is not required.

Results come back as evidence records with confidence, coverage and limitations attached rather than as a confident paragraph. Large results are paged. For an agent, that distinction is the difference between a tool that reports "the function at 0x4a1f20 pushes three arguments and returns a struct pointer" and a model that says something plausible and unverifiable.

So the request that gets refused in a chat window has a different shape here. The model is not doing the analysis. It receives findings from a tool that already looked, and its job is to write code against them. That is also why the thing works headless:

craft-anything-cli commands                        # every tool, by toolset
craft-anything-cli mcp                             # serve them to any MCP client
craft-anything-cli rebuild https://example.com "Recreate this website"

Building does need a coding agent signed in on your machine. Claude Code is supported first: Craft Anything runs your own unmodified claude program, on your own sign-in, and never touches your account. A Codex adapter exists but the README says it has not been verified on a real install. Agents that speak ACP, which includes Gemini CLI, Qwen Code, opencode, goose, Hermes and Maki, can be switched per backend, though only Maki has actually been run live. Treat that list as a roadmap with working wiring rather than a bench of tested options.

Clean Room and Deep Mode

The default is a clean room. Nothing inside the original gets read: not its code, not its files, not its page source. Craft Anything works from what a visitor or user can see, which means screenshots, visible text, measured layout, colours, fonts, spacing, menus you open, and a picture of an app's window if you allow it. Tools that would look inside are refused, and the refusal tells you how to switch to Deep mode.

Deep mode reads everything. Program code and files go through the built-in engines and through Ghidra when it is installed. A website's JavaScript is read, its start page watched, and the builder is allowed to use all of it. You switch modes in the title bar, in Settings, with settings_set deep_mode true, or per CLI run with --deep or --clean-room.

The README is explicit about limits, and this is the part most tools bury. Craft Anything does not bypass DRM, logins or anti-cheat. Rebuilds reproduce 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.

One thing worth saying plainly, because a lot of readers will arrive hoping otherwise: clean room changes what evidence the build has. It does not change the rules that apply to the agent doing the building. If your request is to rebuild a specific commercial game, the Study step will happily inspect it and the Build step is where a refusal can still land, because those two sentences Anthropic added apply to the model writing the code.

How It Compares

Ghidra is the benchmark for free reverse engineering and has been for years, with 97,000 stars on GitHub and a decompiler that professional analysts use daily. It will not build you anything. IDA Free is free but non-commercial, limited to x86 and x86-64, and its decompiler runs in Hex-Rays' cloud, which is a strange place for a binary you were told to keep to yourself. Claude Code on its own is the thing being restricted and has no inspection engine of its own.

Craft Anything is the only one of the four that closes the loop from "how does this work" to "here is a working version", and the only one where a policy change at a model vendor does not take the analysis away from you.

Where Craft Anything Falls Short

The repository was created on 31 October 2026 and the project says early alpha in its own README. That is not modesty, it is accurate. Its status document splits what works, what is partial and what is next, and the honest column is not empty.

• The build step is not free unless you already pay for an agent. Craft Anything costs nothing and is MIT or Apache-2.0. Claude Code needs a subscription, and that is the default builder.
• Alternate agents are mostly unproven. Codex is unverified on a real install. Of the ACP backends, only Maki has been run live. If you are avoiding Claude, you are early.
• 194 open issues in three days. The maintainers are responsive and the release cadence is fast, which is good, and also means you are running alpha software that moves.
• A rebuild is not the original. Layout, style and behaviour get recreated from evidence. Assets, copied text and logos are deliberately left alone, so a rebuild of a media-heavy app or game will look and sound different in ways that matter.
• Deep mode moves the legal weight onto you. The tool tells you what it will read. It does not ask the rights holder.
• Source builds are fussy. Rust 1.97 or newer, plus a list of Wayland and GTK packages on Debian and Ubuntu. The published installers skip all of that, so only compile if you want to.
• The brand is not open. The code is MIT or Apache-2.0. The ArtCraft name, wordmark and logos are trademarks, and forks have to strip them.

Getting Started

The easiest path is the ArtCraft Launcher, currently v0.13.0, which installs Craft Anything and the rest of the suite and keeps them updated. There are also standalone installers: an MSI or portable zip for Windows on x64, arm64 and x86, a signed and notarised universal DMG for macOS, AppImage, Flatpak, deb, rpm and tarball builds for Linux, a FreeBSD tarball, and a WebAssembly viewer that tells you what a file is without the file leaving the browser.

Building from source is one clone and one cargo command:

git clone https://github.com/storytold/craft-anything && cd craft-anything
cargo run --release -p craft-anything                 # desktop app
cargo install --path apps/craft-anything-cli          # CLI as craft-anything-cli
craft-anything-cli doctor                             # checks agents and engines

Projects land in Documents/Craft Anything by default, and each one is a folder rather than a database row: craft.json for the request and history, capture/ for what was recorded, app/ for the generated project, checks/ for each comparison, investigation.cany/ for the evidence and inbox/ for open questions. That layout is why the work stays auditable after the session ends.

Who's Behind It

Craft Anything is built by the ArtCraft team, the same group behind the open-source Crafting Apps. We covered that suite when it was seven Rust rebuilds of Adobe's tools; the site now lists sixteen apps, and Craft Anything is one of them rather than a spin-off side project. The Discord is the community hub, releases are code-signed, and the roadmap is public.

The Bottom Line

Anthropic's rule change is narrow on paper and fuzzy in practice, and the fuzziness is the story: a two-sentence prompt edit is now the gate on work that used to be routine. Craft Anything is a reasonable answer to that, not because it defeats the gate, but because it moves the part that matters out of the model's hands. The engine inspects, records evidence and compares results. The model only writes code.

It is three days old, in alpha, and dependent on a coding agent you probably pay for. If you want to know how a piece of software works, without asking a vendor's permission and without an AI in the reading path, this is the most complete free option that exists right now, and it is worth the patience that alpha software demands.

Craft Anything is also catalogued in our directory.

Looking for more free tools built by small teams in Rust? Read our piece on ArtCraft's seven free Adobe alternatives, our look at Concat, the open-source CapCut alternative, and OpenWorker, a free open-source AI coworker.