Quick Answer: Craft Anything is an open source, pure-Rust engine that studies a website, app, or game with real inspection tools, has a coding agent build your own version, then compares the result with the original and refines it. It's in early alpha, licensed MIT OR Apache-2.0, and runs as a desktop app, a CLI, or an MCP server.
Most people who hear "AI can clone a website" picture a prompt and a prayer. Craft Anything open source takes a different route. It looks at the original first, records what it finds, and only then asks a coding agent to build anything. The project, published by storytold on GitHub, describes itself as a way to understand how any software works and craft your own. Because it's early alpha, the honest version of the story matters as much as the pitch. Here's how the workflow runs, where the trade-offs sit, and what you should check before you trust a build.
The four-step loop: Study, Build, Check, Ready
Most AI rebuild tools start with the generation step. Craft Anything starts with Study, and that order changes the outcome. When an agent gets a vague description, it guesses at spacing, menu behaviour, and layout. Those guesses compound across every page it writes. When the agent gets recorded captures of the original, it works from evidence instead.
The full loop has four stages:
- Study: the engine inspects the target: its pages at every width, its code, its files, and how it runs.
- Build: a coding agent creates a new project in the language, framework, or engine you name, or a sensible default.
- Check: the build is compared with the original, and every mismatch gets listed.
- Ready: you open, play, or preview the result, or ask for changes and the loop runs again.
Study doesn't need AI at all. The engine exposes 284 tools, and each one works as a CLI command and an MCP tool, so you can inspect a target without a language model involved. Only the build step calls a coding agent. That split matters when you're deciding what to trust.
Why does an AI rebuild need a visual check?
Here's the part most guides skip: the check step matters more than the generation step. A coding agent will happily report that its version is finished. Without a comparison, "finished" just means the agent stopped.
For websites, Craft Anything compares pages at three screen widths. The Simple view shows each check as a line item: what doesn't match yet and how alike each page is. That's a far better signal than a single pass or fail. A header that's right at desktop width but breaks at phone width shows up as a specific mismatch, not a vague sense that something's off.
That said, a visual match isn't the same as a functional match. Two pages can look identical while one has a broken form or a dead menu. The check tells you how close the build looks and behaves against what it measured. It doesn't prove the software works the way the original does under every condition. Treat the check as a strong filter, then test the parts you care about yourself.
Clean room reverse engineering or Deep mode: choosing your method
Craft Anything runs in two modes, and the difference is about what the engine is allowed to read. Clean room is the default. It's a form of clean room reverse engineering: nothing inside the original is read, not its code, files, or page source. The engine works only from what a visitor can see: screenshots, visible text, measured layout, colours, fonts, spacing, and menus opened. Tools that would look inside are refused, and the app tells you how to switch on Deep mode.
Deep mode reads far more: program code and files through built-in engines, a website's JavaScript, and the start page's network traffic. It can also call the Ghidra decompiler when that's installed. The README states Deep mode is for your private use, so the usual questions about licensing and permission apply.
| Mode | What the engine can read | Tools available | Best fit |
|---|---|---|---|
| Clean room (default) | Screenshots, visible text, measured layout, colours, fonts, opened menus | Visitor-view tools only; code-inspecting tools are refused | Recreating how a site looks and behaves from the outside |
| Deep | Program code and files, website JavaScript, start-page traffic | All tools, including Ghidra when installed | Studying an app's internals for your own private use |
| Just look inside | The target, for questions only, with no build | Deep mode only | Learning how something works before you decide to build |
Which mode should you pick? Start with clean room. You'll get a faithful visual and behavioural copy of anything a visitor can see, and you won't need to justify reading internals. Switch to Deep only when the visible surface can't explain what you're trying to reproduce, and only for software you have the right to study.
Simple or Advanced: two front doors, one engine
Simple mode is the default and asks, "What do you want to make?" You pick a website, app, or game, describe it, and press Start. The project page then shows the four steps, what the agent is doing in plain words, question and approval cards, and how alike the result is. Ask for changes in the box at the bottom.
Advanced mode is the full workbench. It includes:
- every tool, opened with Ctrl/Cmd+K
- functions, strings, pseudo-C, disassembly, and hex views
- call graphs, the evidence ledger, unknowns, jobs, and logs
- the agent console
Both modes drive the same engine, so nothing is locked behind Advanced. The practical trade-off is visibility. Simple hides the evidence behind each answer, which is fine when a build passes. When a build looks wrong, switch to Advanced and read the evidence the agent cited. You'll find the cause faster there than by re-prompting.
Running a rebuild from the command line
The CLI is where the engine is easiest to script. Three commands cover most cases:
craft-anything-cli commands # every tool, by toolset
craft-anything-cli mcp # serve them over Model Context Protocol (MCP)
craft-anything-cli rebuild https://example.com "Recreate this website"
The rebuild command asks its questions in the terminal. You can also continue an existing project or edit one in place:
craft-anything-cli rebuild https://example.com "Make the header blue" --continue-in "<project folder>"
craft-anything-cli rebuild - "Add a contact form" --project-folder "<your folder>" --existing edit
Each project folder holds craft.json (the request, target, status, and history), capture/ (what was recorded), app/ (the generated project), checks/ (each comparison), investigation.cany/ (evidence), and inbox/ (open questions). Read checks/ first when you review a build. It's the fastest way to see what the agent claims against what it measured.
Results come back as evidence records with confidence, coverage, and limitations. Large ones arrive as a summary, and you page through them with core_evidence_browse. That design keeps an agent's context small, and it also forces you to notice the stated limits.
What the AI requirement means for you
Building needs a coding agent on your own machine. Claude Code is the supported builder today. Craft Anything runs your own, unmodified claude program, which uses your own sign-in. The project says it never sees or handles your account, and your use falls under your own agreement with Anthropic. The repo's docs/claude-code.md explains the details.
Other options are less proven. A Codex adapter exists, but the README says it hasn't been verified on a real install. Anthropic, OpenAI, OpenRouter, and local OpenAI-compatible servers work for questions and analysis under Settings → AI. Builds using API keys are implemented but haven't been run live. Ask yourself which path you'd trust before you commit a project to it.
Where Craft Anything is still rough
The README calls Craft Anything early alpha, and the status breakdown is worth reading closely:
- Works today: the desktop app in Simple and Advanced, the CLI and MCP server, the Study → Build → Check → Refine loop for websites, apps, and games using Claude Code, website checks at three widths, game rebuilds in Unity, Unreal, or Godot, and Deep mode with Ghidra.
- Partly there: behaviour checks for apps and games, builds through API keys and other agents, and some engine tools tested only against stand-ins.
- Next: the first public release, with installers for Windows, macOS, Linux, and FreeBSD, plus a browser viewer, both set up in CI.
You'll also need Rust 1.97 or newer to build from source. Most of the limits come down to one thing: the check is only as strong as the behaviour it measures. For a landing page, that's usually enough. For a game with complex logic or an app with real data, plan to test the parts a visual comparison can't cover.
Frequently Asked Questions
Is Craft Anything open source?
Yes. Craft Anything is an open source project under the MIT OR Apache-2.0 licenses, written in pure Rust. It's in early alpha, so expect rough edges. You can build it from the GitHub repository with cargo run --release -p craft-anything for the desktop app.
How do I rebuild a website with a coding agent using Craft Anything?
Install the CLI with cargo install --path apps/craft-anything-cli, make sure Claude Code is signed in, then run craft-anything-cli rebuild <url> "Recreate this website". Use clean room mode by default. When it finishes, read the checks folder to see which pages and widths still differ from the original.
When should I use Deep mode instead of clean room?
Use Deep mode only when the visible surface can't explain what you need to reproduce, such as a program's internal logic. It reads code, files, and JavaScript, and it can use Ghidra. Use it only on software you have the right to study, since the README says Deep mode is for private use.
Does Craft Anything need an AI model to study software?
No. Inspection runs on the engine's own tools, which work without a language model. You need a coding agent only for the build step, and Claude Code is the supported option today. craft-anything-cli doctor checks your install and tells you which agents and engines are missing.
Craft Anything open source is most useful when you treat the check list as the real output, because a build is only as trustworthy as the comparison behind it. Install the CLI, run craft-anything-cli doctor, then rebuild a simple page you own and read every mismatch it reports. Pass this along to someone who keeps asking whether AI can copy software without learning how it works.