Personal project · 2026 · In progress

Craft AnythingReverse engineering · Local background removal

Discovering how an AI agent can study existing software, then build its own version of it.

Craft Anything is an open-source reverse-engineering toolkit. To learn how to use it, along with the ArtCraft suite, I ran it on remove.bg, an online image background removal service. Under my direction, AI agents then rebuilt its features in a Rust program that I use offline.

RustONNX RuntimeWebAssemblyBiRefNetAI
Alpha channel

Try it

Remove a background without sending anything

The processing runs in your browser: your images never leave your device. This online version uses a lighter model than the program, so results may be less accurate.

Demo

On the first run, the browser loads the engine once, about 60 MB, so Wi-Fi is recommended. On an older phone, processing may fail for lack of memory.

The tool

How Craft Anything works

Craft Anything does more than look at the screen. Its Rust engine offers over 300 inspection tools, available from the command line or through the MCP protocol, which a coding agent such as Claude Code calls directly to analyse the target.

  • clean room

    The default mode

    Nothing inside the original is read: no code, no files, no page source. The analysis starts from what a visitor sees, at several screen widths: screenshots, visible text, measured layout, colours, fonts and opened menus.

  • deep

    The advanced mode

    Meant for private use, it opens up every tool: Mach-O structure, Objective-C classes and demangled Swift names, decompilation with ILSpy for .NET and Vineflower for Java, JavaScript mapping with js_app_map, and watching a running program with Frida.

  • rebuild_check

    Build, then compare

    The results are organised into call graphs and component lists, from which the agent draws up a build plan. It then compares its version with the original, notes what does not match yet and fixes it.

Method

From observation to tool

  1. 01

    Observe

    Capturing the service as a visitor sees it, at three screen widths: pages, flows and behaviour. Its HTML, scripts and requests were not reused.

  2. 02

    Isolate the features

    Drag and drop, batch processing, input formats and transparent export: the list of what the program must be able to do.

  3. 03

    Rebuild locally

    A Rust program, a native window and a segmentation model running on the machine, with no remote service.

  4. 04

    Verify

    End-to-end tests on the real program and scoring of cutouts on images whose expected result is known.

Processing chain

From file to transparent PNG

  1. ImageDecoding, orientation and scaling.
  2. Colour analysisOn a logo or diagram with a plain backdrop, AI is not even needed.
  3. ONNX modelBiRefNet in the program, silueta on this page.
  4. Rust refinementMask and edge correction on light backdrops.
  5. Alpha PNGOnly the subject remains, the background becomes transparent.

Architecture

Three Rust building blocks

  • sansfond

    Application

    The window, embedded pages and cutout engine, combined in a single program. It accepts PNG, JPEG and WebP images, one at a time or in batches.

  • modelpack

    Compression

    A small palette of values per weight table takes the model from 109 to 40 MB.

  • refine

    Post-processing

    Light-backdrop refinement, compiled natively for the program and to WebAssembly for this page.

Project status

Where the program stands

  • WindowsTested
  • Linux 64-bitTested
  • macOS, Apple chipsBuilt, still to be tried