Opus 5.5-Built Photoshop Clone PhotoCraft Hits 41K Stars and Rates Itself 45% Ready

TL;DR

ArtCraft founder Brandon Thomas pushed seven clean-room, pure-Rust reimplementations of Adobe's creative suite to GitHub between September 30 and October 1, built by Claude Opus 5.5 agents running in Claude Code. The Photoshop stand-in, PhotoCraft, went from 6,691 stars on October 7 to 41,572 on October 10, and the seven Adobe repos together sit just under 80,000. The same day, the project published a full re-measure of itself against Adobe Photoshop 2026: roughly 76% feature breadth, but about 45% "ready for real work," 3 of 25 performance budgets met, and 1,100 to 1,800 Opus 5.5 agent-hours still to go before beta. Its own October 3 estimate of 140 to 220 hours to 95% parity is now labeled "far too optimistic." That written retraction is the most useful artifact in the story.

What shipped

The storytold org holds the lot: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere Pro), LightCraft (Lightroom), PdfCraft (Acrobat, briefly called PrintCraft), EffectCraft (After Effects) and DesignCraft (InDesign). All seven are Apache-2.0 by GitHub's detection, with PhotoCraft's README offering MIT or Apache-2.0. Native builds target macOS, Windows and Linux, and six of the seven also compile to WebAssembly.

The stack is deliberately narrow. The AGENTS.md every agent reads first bans Electron, Tauri and any JavaScript UI: the desktop app is egui on wgpu, the web build is the same Rust compiled to WebAssembly, and GitHub counts about 18.7 MB of Rust source in the PhotoCraft repo alone.

The design decision worth stealing is the command registry. Every menu item, tool and dialog dispatches one of 500-plus commands, and the UI, the CLI, a JSON control channel and an MCP server all call the same registry. If agents can drive the app at the command level from day one, the agents writing it can also test it, which is much of how a ten-day-old codebase got this far.

Thomas has not stopped at Adobe. On October 7 the org added WordCraft (Word), GridCraft (Excel), DeckCraft (PowerPoint), CADCraft (AutoCAD-style drafting) and SoundCraft (Pro Tools), much as he promised on Hacker News two days earlier, and a launcher that installs all of them landed on October 9.

Eleven days of commits

Thomas said the quiet part on Hacker News on October 2: "I'm currently using Opus 5.5 to create 100% pure-Rust, immediate mode rendering, open source implementations of Adobe's entire suite." On October 10 he added that the team runs "vanilla Claude Code" and is "especially careful not to touch Ghidra," since decompiling Adobe binaries would void the clean-room claim. Elsewhere on HN he says he stopped writing code by hand in February.

The repo's own calibration section quantifies the throughput: 1,152 commits on main and 1,004 merged PRs in the 10.5 days to October 10, a median PR open-to-merge time of 3.4 hours, and "the repository has sustained 10 to 20 concurrent agents." The release page shows eight tagged builds between October 2 and 10, ending at v0.6.0. At the project's rule of thumb of two to four agent-hours per merged PR, that is 2,000 to 4,000 Opus 5.5 agent-hours spent. The parallel-agent rules are mundane and copyable: a private Cargo target directory per agent, new commands in new modules rather than shared files, and manifests written to a temp file then moved into place so one half-written Cargo.toml never breaks everyone's build.

The attention arrived early. Someone else posted the apps page to HN on October 4, where it drew 223 comments, and a second thread on the repo added 77 more on October 6. Thomas's reply: "99% parity will take a while, but I'm sure it'll be measured in months and not years. Feel free to check in on this claim. It's measurable." That was already a retreat from the Reddit launch posts, which promised 100% parity within a month. The one-month promise lasted about three days.

The self-audit

What makes this more than another vibe-coded clone is that the project checked in on its own claim, in writing, on October 10. The new ROADMAP.md and target-app parity document re-measure PhotoCraft 0.6.0 against an installed Photoshop 2026 (27.11.0), read from the bundle's shortcut file, plug-in lists and binary strings without ever launching it. The clean-room line held even during measurement.

The result is two numbers pulling in opposite directions. Feature breadth lands around 76%: 628 of 628 catalogued menu commands dispatch a live command, and 53 of 68 toolbox tools and 30 of 35 panels exist. Ready for real work, weighted by what a working retoucher or designer depends on, lands around 45%, with a stated range of 40 to 50%. Breadth asks whether the key exists on the piano. Readiness asks whether it is in tune, survives a two-hour set, and sounds like the instrument you learned on.

The document is candid about why the first estimate was wrong: "about 1,230 issues have been filed since the first release (579 open on 2026-10-10), most of them against features already counted as present." An interim "honest assessment" on October 5 had put readiness at 25 to 35%; roughly 700 merged PRs since then moved it to 45%.

By dimension, the shape of the gap is clear. Platforms are at 80%, but performance sits at 25%, the plug-in and scripting ecosystem at 15%, and AI features at 10%, because there is no ML runtime in the app at all, so no Generative Fill, Neural Filters or Sky Replacement.

Where it actually breaks

The generated scorecard says what "25% on performance" means in milliseconds. On an Apple M4 Pro baseline from October 5, changing a layer's opacity has a budget of 20 ms at p95 and measured 443 ms. Pasting a 12-megapixel layer has a budget of 100 ms and measured 3,136 ms. An arrow-key nudge in a 150-layer document panics the GPU path. Content-Aware Scale on a 24-megapixel image has a 3-second budget and came in at 954 seconds, which is less a performance gap than a coffee break.

PSD fidelity is the other gate. Inside the project's own corpus, round trips are near perfect (308 of 309 psd-tools files), but against Photoshop-authored oracle files the composite matches only 133 of 256, and the gaps document cites user reports of Photoshop refusing to open PhotoCraft's exports. It is blunt that the only real oracle is "opening exports in Photoshop on the owner's Mac," and marks that, the AI-backend decision and Windows code signing as work an agent cannot finish alone. Fifteen of Photoshop's 68 tools are still missing, and Select and Mask, Vanishing Point and Content-Aware Fill exist as parameter dialogs rather than workspaces.

HN testers found the same edges by hand: one could not paste from the clipboard, one found the Flatpak files the README references were never shipped, and a photographer called LightCraft's raw output "an amateur attempt at RAW processing from first principles." The maintainer of the printpdf crate warned that correct PDF font subsetting "took months of trial and error." One commenter's verdict on the whole venture, "Icarus with a claude jetpack," at least fits on a sticker.

The economics and the lawyers

The apps are free, carry no telemetry and need no account. Adobe's current US list, per its own plans page, is US$22.99 a month for a single app such as Photoshop and US$69.99 for the all-apps Pro plan on an annual commitment; Thomas's stated motivation is a cancellation fee he hit "one too many times." Per Gadget Pilipinas, the work is funded by selling token-based access to ArtCraft's existing hosted generation models. Thomas says he will pay stipends and bug bounties, and the contributor list already shows a dozen-plus community accounts behind his own.

Legally, the README states the apps are "implemented from public specs and observed behaviour only. No proprietary code, shaders or assets," and Ghidra is banned. What critics raise is trade dress: menus, shortcuts and panel names are intentionally identical, and FilmCraft keeps the "Lumetri Color" panel name. Tedium's Ernie Smith, broadly sympathetic, still wrote: "Who knows how legal this is? This feels like one a legit lawyer should probably weigh in on." Adobe has not commented in any coverage we found.

How much is left, in agent-hours

The most reusable idea in the repo is that it prices its backlog in a measurable unit. Using merged PRs as calibration (a Remove tool cost 9 PR-hours and 2,005 lines; a Red Eye tool 17.9 hours), the project estimates 2 to 9 agent-hours per checklist item and 8 to 20 per new format or subsystem. Summed across the gaps list, that is 1,100 to 1,800 Opus 5.5 agent-hours to beta and 2,300 to 4,100 to full parity. Since about 80% of the work parallelizes, beta is "roughly 100 to 150 wall-clock hours" at a dozen concurrent agents.

Whether you believe the hours is almost beside the point. A parity claim became a document with weights, evidence and a method, backed by a scorecard that a cargo task regenerates and CI checks, and it publicly corrected its own author one week in. Most agent-run projects produce a README that says "100% parity" and a closed issue tracker.

What to take from it

Three things transfer to anyone building with a fleet of coding agents. Make every user action a command in one registry before you build the UI, so agents and tests drive the same surface humans do. Keep a measured readiness doc that separates "exists" from "works like the real thing," because agents are superb at the first and your issue tracker fills with the second. And calibrate estimates in agent-hours from your own merged PRs, then publish the ranges. If you are a Photoshop user wondering whether to switch, the project's own number for a typical pro is 43%: wait for the beta and read the PSD line of the scorecard first.

Key Takeaways

• ArtCraft's seven clean-room Rust rewrites of Adobe's suite, built with Claude Opus 5.5 agents in Claude Code, hit about 80,000 combined GitHub stars in ten days; PhotoCraft alone went from 6,691 to 41,572 between October 7 and 10.
• Velocity by the repo's own calibration: 1,152 commits and 1,004 merged PRs in 10.5 days, a 3.4-hour median merge time, and 2,000 to 4,000 agent-hours spent.
• The October 10 re-measure against Photoshop 2026 27.11.0 reports 76% feature breadth but 45% "ready for real work," and calls the October 3 estimate of 82% and 140 to 220 hours to 95% "far too optimistic."
• The weak spots are concrete: 3 of 25 performance budgets met, 133 of 256 Photoshop-authored PSDs rendering correctly, 15 missing tools, no ML runtime.
• Remaining work is priced at 1,100 to 1,800 Opus 5.5 agent-hours to beta and 2,300 to 4,100 to full parity, with PSD verification, the AI backend and code signing flagged as needing a human.
• Free, Apache-2.0 or MIT, no telemetry, funded by ArtCraft's hosted generation tokens; the open legal question is trade dress, not copied code.

Sources: storytold/photocraft on GitHub, PhotoCraft target-app parity document, PhotoCraft ROADMAP.md, PhotoCraft scorecard, PhotoCraft gaps document, PhotoCraft AGENTS.md, storytold GitHub organization, ArtCraft apps page, Hacker News thread, October 4, Hacker News thread, October 6, Brandon Thomas on HN, October 2, Brandon Thomas on HN, October 10, Adobe Creative Cloud plans, Gadget Pilipinas, HotHardware, Notebookcheck, Tedium