When a project grows this fast, the marketing and the issue tracker usually tell two different stories. craft-anything is worth a second piece because it does not pretend otherwise.

Its status section is blunt: early alpha. What officially works today is the desktop app, the CLI, the MCP server, and website rebuilds with Claude Code as the builder. What is partly there includes building through API keys and other coding agents, plus behaviour checks for apps and games. Generated images and 3D models are opt-in and need your own fal.ai key.

The issue tracker turns those three tiers into specific scenes.

First failure: the builder is not Claude Code

In issue 736, a reporter used gemini-cli as the builder and hit a loop: the checker kept saying no built site existed in app/dist, the agent insisted it had run npm run build, and app was empty. The checker was right to refuse; the loop itself is the bug.

Issue 478 is blunter. On v0.3.0, Claude Code builds failed after two seconds because a resume call passed an invalid session id. The same user reported Codex worked.

Issue 498 concerns a local model: Qwen3 8B sent an array where the core_unknown_add tool expected a string. Stricter argument parsing is now listed in the status file.

All three point the same way. Only the Claude Code path has been exercised repeatedly. The others run, but their edges have not been mapped.

Second failure: clean room is a policy, not a sandbox

This is the most misread part, and the project says it repeatedly.

Clean room is a set of rules enforced at the tool layer. It does not isolate your machine and does not restrict the coding agent, which still runs as you; what it can read and write depends on its own configuration. Not reading source is a refusal by the tool, not an interception by the operating system.

The companion default is worse for the unwary: Auto mode is on by default, so commands do not each prompt you. The docs suggest leaving the package manager open only for projects whose dependencies you trust, and letting other commands go through the approval card.

A third risk comes from outside reviewers: training-data contamination. If the agent's training corpus includes the target's code, the clean-room defence weakens sharply. And the derivative-work question for AI output is still unresolved. Interoperability research sits on solid ground; cloning a competitor and shipping it commercially is a different calculation.

Third failure: a good score is not a working product

The Check score measures appearance.

Reviewers put it plainly: a page at 90 percent visual similarity can still have broken forms. The number is useful for judging how far you are, not whether you can ship. It is also why behaviour checks for apps and games are listed as partly there.

A few practical numbers. v0.5.0 downloads range from about 47 MB to 133 MB, the macOS DMG being largest. Deep mode with Ghidra needs Java and can take up to half of system memory. Building from source needs Rust 1.97 or newer, and on Linux a stack of X11, Wayland, GTK and OpenGL development libraries.

Release velocity cuts both ways

Six versions in ten days is a moving target. Good, and a warning that the interface and docs you read this week may change next week.

The tool count in the README moved from 301 to 314 to 320 within a fortnight, while independent reviews counted about 300 inside the reverse-engineering engine and the rest in the AI and setup layers. That drift is normal in early projects, but cite a version when you quote a number.

What to do with this

Start with Claude Code, a small website you own, and one full pass through the four steps, then decide whether to wire in other agents. Do not make your first project a commercial clone; that crosses more than a technical line.

On the legal side, settle three questions first: do you have the right to examine the target, will the output carry the other party's marks and copy, and is this for internal use or publication. The tool can only help with part of the second.