AI News

ArtCraft open-sources Craft Anything: research the target first, then have the Agent imitate it

Shao MengUpdated Nov 7 21:46

AI summary

ArtCraft, which previously open-sourced a full suite of Photoshop and Office tools, has open-sourced Craft Anything, letting the Agent first collect information about the target's appearance, layout, and behavior, and then build its own version based on that; the web page inspection covers three screen widths. The project provides more than 300 tools and supports both a command line and an MCP interface, so it can serve as a research toolbox for other Agents. The default clean-room mode uses only external information such as screenshots, visible text, and layout, and sets a tool whitelist, while the deep mode can analyze code, files, and web page scripts; currently the build path calls the locally logged-in Claude Code, with costs borne by the user, and image and 3D generation additionally require a fal key. The project uses a dual MIT / Apache-2.0 license, and adaptations such as Codex have not yet completed real-environment verification.

Why read this

AI score 62

The original text explains the boundaries of the clean-room and deep research modes, as well as the current state of building depending on the local Claude Code, making it easier to judge the scope of usability.

Topics

Tags

Body · AI translation

Remember @get_artcraft, the folks who open-sourced Photoshop and then the whole Office suite? They've now open-sourced their signature reverse-engineering expertise!

Craft Anything: research first, then build, so AI imitation has a basis and can be checked https://t.co/nKgS8B6bN0

Not just describing requirements, but giving the Agent a basis

After the user specifies a target, the tool first collects information such as appearance, layout, and behavior, then hands it to a coding Agent to build its own version. Web page inspection covers three screen widths, helping spot differences between the finished product and the reference object, rather than ending once the code is written.

The project provides more than 300 tools and supports both command-line and MCP interfaces, so it can also serve as a research toolbox for other Agents instead of being used only through its interface.

Two research modes, distinguishing appearance reference from internal analysis

The default "clean room mode" uses only externally visible information, such as screenshots, visible text, and layout. The code includes a tool whitelist that restricts reading the target's internal implementation.

"Deep mode" can analyze code, files, and web scripts. The difference between the two modes is that it separates "referencing how a product behaves" from "studying internal implementation"; however, this technical distinction does not automatically resolve copyright or licensing issues.

What the build depends on, and how mature it is

The currently defined build path is to call the locally logged-in Claude Code, with costs borne by the user; image and 3D generation additionally require a @fal key. The project uses a dual MIT / Apache-2.0 license.

Although there is adaptation code for other Agents such as Codex, the project documentation still notes that real-environment verification has not been completed, so you can try it yourself.

Source

Shao Meng