Kefong Studio 科奉工作室

AI Rewrites the Adobe Suite and Office: Clean-Room Clones — Progress, Controversy, and the One Change That Really Matters

·6 min read

Written in October 2026. Data comes from the project repositories and public documentation (star counts and completion figures are snapshots from around 2026-10-12 and will change over time).

In one sentence: in late September 2026, a developer used AI coding tools to ship 7 open-source apps taking on the Adobe suite within a single week (pure Rust, clean-room implementations), then turned to Office; they picked up tens of thousands of GitHub stars in a week. But the project's own completion documents are remarkably honest — between "it runs" and "it's usable" lie over a thousand agent hours. What really deserves attention is their architecture: every action is a named command, exposed simultaneously via CLI, a JSON channel, and MCP, so the apps are natively drivable by AI agents.

What happened

The central figure is Brandon Thomas, and the projects live under the GitHub organization storytold, collectively called "Crafting Apps." The development tool is Anthropic's Claude Opus 5.5, used with multiple agents in parallel (the author says 8 parallel agents), all written from scratch in Rust.

Timeline (from the progress log in the PhotoCraft repository):

Date Progress
2026-09-30 First commit
2026-10-01 Menu coverage went from 224 to 532 of 625 in a single day
2026-10-03 Menu breadth 625/625
2026-10-05 0.2.0, the first signed/notarized release; self-assessed as "able to do real work 25–35%"
2026-10-07 0.3.0, 10 UI languages, keyboard shortcut audit went from 214 → 0 failures, TIFF decoding 10× faster
2026-10-08/09 0.5.0, added Affinity document import and a batch of tools
2026-10-10 Full re-measurement against Photoshop 2026, 0.6.0

The approach: a clean-room rewrite

Clean-room means: never looking at or copying the other party's source code, and reimplementing equivalent functionality based only on public documentation + observable external behavior. The repositories contain no Adobe code, icons, or presets, and the contribution rules explicitly forbid copying.

  • For the framework, everything is native Rust with a custom-built GPU compositor (Metal / Vulkan / D3D12) — no Electron;
  • File formats are implemented in-house: PhotoCraft has a standalone photocraft-psd for PSD/PSB, claimed to render identically after re-saving on 307/309 files in the psd-tools test suite; PhotoCraft is about 256,000 lines of Rust;
  • Licensing is dual MIT / Apache-2.0, free to fork.

How much is done (by the project's own accounting)

That level of honesty itself became the focal point of discussion online. PhotoCraft's ROADMAP.md scores itself directly against Photoshop 2026 (27.11.0, measured against a locally installed copy):

Metric Value
Feature breadth ~76% (menus 628/628, tools 56/68, panels 30/35, formats 13+4 / 31)
Able to do real work ~45%
Replaceable for mainstream professional users ~43%
Users who only use basic features ~61%
Effort still needed to reach beta ~1,100–1,800 agent hours
To full parity ~2,300–4,100 agent hours (about 80% parallelizable)
Current stage alpha

Broken down by dimension, the gaps are clearer:

Dimension Readiness
Platform support 80%
File formats 60%
Feature depth 55%
Localization 55%
Hardware (GPU / drawing tablets) 50%
Stability 45%
UI/UX fidelity 40%
Performance 25% (only 3 of 25 budgets met)
Plugin / scripting ecosystem 15%
AI generation features 10%

Two more very concrete signals: the repository has 579 open issues; and PSDs it writes currently can't be reliably opened by Photoshop (see issues #1281 and #2469). In other words, "307/309 identical after re-save" is measured on its own test suite; crossing over to Adobe's actual product still breaks.

The lineup (star snapshot 2026-10-12)

App Takes on stars
PhotoCraft Photoshop 53.1k
FilmCraft Premiere Pro 11.4k
LightCraft Lightroom 11.2k
PdfCraft Acrobat 11.2k
VectorCraft Illustrator 7.8k
EffectCraft After Effects 5.8k
DesignCraft InDesign 3.4k

Then came Office

The same organization went on to apply the method to the office suite, with the same approach (pure Rust, clean-room):

App Takes on
WordCraft Word (reads/writes .docx, ribbon / styles / tables / track changes / mail merge)
GridCraft Excel
DeckCraft PowerPoint (2.6k stars)

But this space was actually already crowded with "AI-native office" tools, which are worth looking at together:

Project Highlights stars
OfficeCLI Single-file binary built specifically for agents to read/write docx/xlsx/pptx; 350+ Excel functions computed on write; can render documents to HTML/PNG, closing the "render → look → fix" loop 30.6k
GenOffice Full suite (Docs/Sheets/Slides/PDF/HTML/Markdown), byte-level preservation of unmodified content, ships with a CLI, an agent skill, and a local MCP 8.2k
HermesOffice README is strikingly similar to GenOffice's (down to the BYOK list and the Genspark login method); looks like the same suite forked under a new name 584
WPS-AI A WPS plugin; the author states the repository contains "not a single line of hand-typed code" —

And one phenomenon worth pondering along the way: the barrier to cloning is now so low that "change the name and it's a new product."

The genuinely new thing: commands + CLI + MCP

These projects share one architectural trait, and it's the one most worth remembering:

  • Every menu item or tool you click in the UI calls the same named command behind the scenes (PhotoCraft has 500+);
  • The same set of commands is exposed simultaneously to the GUI, the command line, a JSON control channel, and an MCP server;
  • The repositories also include instructions written for AI to read (EffectCraft's agents.md, OfficeCLI's SKILL.md), so an agent can read them, set up the environment on its own, and get to work.

The result: "batch-edit 300 images / 300 documents with a single prompt" goes from impossible to an engineering problem. Traditional desktop software locks its capabilities inside the GUI, while this generation of alternatives turns those capabilities into interfaces that programs (and AI) can call. That's where the impact is — not in "free."

Controversy and risks

  • It runs ≠ it's usable: building it in 48 hours is real, and needing 1,000+ more hours before it's safe for production is also real. The project's own ROADMAP happens to be a living example of The Mythical Man-Month — AI hasn't eliminated the cost of coordination and verification, it has just changed the unit from "man-months" to "agent hours."
  • Why not contribute to existing open source? The author has publicly slammed GIMP as "terrible" and complained that LibreOffice wrecked his final-term paper, which drew criticism that "rather than rewriting, the effort would be better spent on mature projects" (peers also pointed out that LibreOffice and Google Docs have long been available for free).
  • Legal boundaries: clean-room implementation of functionality generally holds up, but close similarity of the interface, icons, and menu structure falls into the gray zone of trademark and "look and feel" — there's no case law yet.
  • The ecosystem gap is a hard number: plugin ecosystem 15%, AI features 10%. A large part of Photoshop's value lies in its plugins and the .psd asset chain, and that's not something you can close just by "writing the code."

How to try it yourself

  • Traditional desktop apps (the Craft series): GitHub Releases has Windows / macOS builds, and Linux can be built from source; there's also a WebAssembly browser version (according to the project's notes). Licensed MIT / Apache-2.0.
  • Agent-friendly office tools: installing OfficeCLI once writes its skill into the coding agents it detects on your machine (Claude Code / Cursor / Windsurf, etc.); GenOffice ships with the genoffice command line and an agent skill, and the desktop app is free for standalone use.
  • Advice: treat them as a testing ground; don't uninstall your licensed software. Use them to try "batch, scriptable, agent-driven" tasks, and keep using the original tools for real deliverables.

Conclusion

  • This is a demonstration of a methodology: clean-room + multiple agents in parallel compresses "software that takes a big team years to build" into "one person, first version in a week." The first version runs, but there are still 1,000+ hours of hard work between it and "shippable."
  • The more practical impact on the industry is the architectural legacy: a command registry + CLI + JSON + MCP, which for the first time lets AI drive design, video, and document tools headlessly in batch.
  • For users, the call is simple: for productivity, stick with the original tools for now; for automation, you can start trying these alternatives today.

References