Someone Rebuilt Adobe in Rust with AI Agents. Is It Usable?

Last Updated: October 8, 2026

Key takeaways:

A developer named Brandon Thomas published clean-room Rust rewrites of seven Adobe apps under the ArtCraft brand on September 30 and October 1, 2026, built with Anthropic's Claude Opus 5.5 coding agents running about 8 in parallel.

PhotoCraft, the Photoshop clone, hit 19.4K GitHub stars within 8 days, and five more clones followed targeting Word, Excel, PowerPoint, Pro Tools and AutoCAD.

The parity claims need context: 100 percent menu coverage means every menu item dispatches a command, but the project's own assessment puts professional daily-work parity at 25 to 35 percent, with 68 percent of test PSD files rendering pixel-identical.

The load-bearing part of the workflow is a test oracle: around 1,700 tests plus a composite oracle that diffs PhotoCraft's render against Photoshop's own merged output decide whether agent-generated code counts as real progress.

The apps are free, offline, cross-platform and genuinely impressive, but the project itself says not to use them for professional daily work yet. Adobe alternatives like Affinity and DaVinci Resolve remain the safer production bet.

What happened: seven Adobe apps rebuilt in Rust in a week

On September 30, 2026, a GitHub account called storytold published PhotoCraft, VectorCraft, FilmCraft, LightCraft and PdfCraft, five apps that reimplement Photoshop, Illustrator, Premiere Pro, Lightroom and Acrobat Pro. On October 1 it added EffectCraft for After Effects and DesignCraft for InDesign. A week later the same account dropped WordCraft, DeckCraft, GridCraft, SoundCraft and CadCraft, taking aim at Microsoft Office and AutoCAD. Every app is pure Rust, MIT or Apache-2.0 licensed, runs on macOS, Windows and Linux plus the browser via WebAssembly, and carries no telemetry. According to GitHub's API on October 8, PhotoCraft alone had 19,456 stars and the 13 clone repos held 41,000 combined stars, all earned in about 8 days.

The developer, Brandon Thomas, is a former Square engineer with about a decade of Rust experience who says he has coded exclusively with AI tools since February. He announced the suite on Reddit and Hacker News, credited Anthropic's Claude Opus 5.5 for the pace, and described the apps as "super early alpha." His now-viral framing: "Software is over." The number that made the story travel is engagement, not just star count. A tweet by Kevin Picchi summarizing the repos did 1.9K likes and was screenshotted into a TikTok post with 20.5K likes, which is where most people first met the story.

Is PhotoCraft really 100 percent parity with Photoshop?

No, and the project's own documents explain why that impression is wrong. According to the PhotoCraft repository's parity estimate, updated October 3, all 625 Photoshop menu items dispatch a live command, which their tooling reports as 100 percent menu breadth. The same document weights depth and fidelity and lands at roughly 86 percent functional parity for mainstream workflows, and about 75 percent once Adobe's AI and Firefly features are included. The honest parity assessment of October 5 goes further: "real Photoshop parity is still well below 50 percent," and the share of what a professional touches that exists in usable form is estimated at 25 to 35 percent.

The fidelity numbers are measured, not vibes. PhotoCraft opens and re-saves 307 of 309 files in the psd-tools test corpus identically, and its composite oracle renders 115 of 170 reference files pixel-identical to Photoshop's own merged output, about 68 percent. About 20 Photoshop tools are still missing, including Patch, Content-Aware Move, Red Eye and the Freeform Pen. The generative AI tier, the feature most consumers associate with Photoshop today, sits at roughly 0 percent and is deferred by an explicit project decision. VectorCraft reports 69 to 75 percent of Illustrator's features, and PdfCraft says about half of Acrobat Pro's offline features are in. To the project's credit, all of this is written in their own roadmaps, which is more honesty than most press releases about AI software ever manage.

How does one developer build seven desktop apps in a week?

He does not, not alone, and not from nothing. The workflow visible in the repositories has three parts. First, a fleet of parallel agents: PhotoCraft's planning documents schedule remaining work in "agent-hours" and record that one session "landed ~70 menu items + major subsystems in a handful of wall-clock hours using 8 parallel agents," with observed parallelism of 6 to 8 times across independent subsystems. Each agent owns a target directory: the PSD parser, the GPU compositor, the text engine, the filter set, the MCP server.

Second, a clean-room constraint. The team forbids Adobe code, shaders, ICC color profiles, fonts and trained models. Everything is reimplemented from public specifications like Adobe's published PSD format documentation and from observed behavior, which is why some raw formats are permanently blocked: Nikon's compressed NEF and Canon CR3 cannot be decoded clean-room because their specifications are not public.

Third, and most important, a test oracle that decides what counts. More than 1,700 tests cover PSD round trips, a CPU compositor that serves as a reference for the GPU compositor, and the composite oracle that diffs output against Photoshop's own rendering. The oracle is why the honest numbers exist at all: agent fleets generate breadth fast, and breadth is exactly what shallow metrics reward. A test that measures against ground truth is what stops 100 percent menu wiring from becoming the headline.

What does this cost compared to Adobe Creative Cloud?

The apps are free to download, with no subscriptions, no accounts and no cloud requirement. Adobe's own plans page lists Creative Cloud Pro covering 20-plus apps at 69.99 US dollars per month on an annual plan, with single apps at 22.99 dollars per month each. Buying the seven equivalents a la carte runs about 147 dollars a month. The ArtCraft projects, including PdfCraft and LightCraft, cost nothing.

The build side has a published price too. According to the project's own estimate, reaching 95 percent functional parity takes 140 to 220 Opus 5.5 agent-hours, which is 25 to 40 wall-clock hours with 8 parallel agents. Anthropic lists Opus 5.5 at 4 dollars per million input tokens and 20 dollars per million output tokens, with cache reads at 0.20 dollars per million, and cache reads dominate long agentic sessions. The model bill for the entire suite is a rounding error next to one year of Creative Cloud for a single seat, which reframes the story: the expensive part of software was never the code, it was the market power. That power is what just got repriced.

Funding for ongoing work comes from Thomas's pre-existing ArtCraft visual AI studio, which sells token-based access, rather than from the clone apps themselves. He has said the suite will not be an orphaned small-team project and is recruiting contributors with stipends and bug bounties.

Is this legal? The clean-room question and trade dress risk

Two separate legal questions are stacked here, and only one is answered. The method is the settled part: clean-room reverse engineering, rebuilding functionality from observed behavior and public specs without copying code, has generally been found legally acceptable, according to the Electronic Frontier Foundation's reverse engineering guidance. The ArtCraft apps follow the strict version of this: no proprietary code, no shaders, no Adobe ICC profiles, no trained models.

The unsettled part is trade dress. Ars Technica's coverage notes that UIs deliberately mimicking Adobe's familiar appearance could still create legal exposure even when the underlying code is clean. Adobe's trademark notice inside the READMEs shows the team is aware: the apps are careful to describe themselves as compatible with Adobe workflows, not affiliated with Adobe. Whether menus, panels and shortcuts arranged the Adobe way constitute protectable trade dress is a question nobody has answered in court, and the answer shapes every future clean-room clone of commercial software.

Should you use the ArtCraft apps today?

For professional client work, no, and the project itself says so. The PhotoCraft status banner states plainly that it is "not yet a Photoshop replacement for daily professional work," and early testers found broken basics that its test suite had passed, including 214 keyboard shortcut failures and an immovable crop frame, all fixed within days but emblematic of alpha software. Most fixes since the 0.2.0 release have not yet been validated by real users.

For everything else, the calculus is different. The apps are worth an hour of any developer's time: they build with cargo, run headless through a CLI, expose MCP servers for agent control, and open real PSD files, which makes PhotoCraft useful today as a second opinion on a file. Linux users gain the most, because Adobe sells nothing for Linux and these apps are native there. Anyone on a locked-down budget should test on copies of files, keep Affinity, DaVinci Resolve and the mature open-source tools for production, and watch the project's contributor count, which at 66 and climbing is the clearest signal of whether this becomes infrastructure or a footnote.

In our own dual DGX Spark setup running open-weight models, the pattern is familiar: agent fleets produce breadth at a rate that shocks people and depth at a rate that does not. The ArtCraft repos are the cleanest public demonstration yet of both halves of that sentence, and the test-oracle discipline they apply to AI-generated code is the part worth copying regardless of what happens to Adobe's market share.

The bottom line on the Rust Adobe clones

The true one-line summary is printed on the repo itself: early alpha, and it says so. What matters more is what the project proves about the economics of software. One engineer with 8 parallel agents, a clean-room constraint and a serious test harness produced 13 desktop applications and 41,000 stars in 8 days. The floor price of a competent creative application is now zero for anyone willing to run early software and file bug reports. Adobe's 23.77 billion dollar revenue base does not shake this quarter, but the argument for paying that base got harder to make in public, and the next incumbent watching this unfold should assume the same playbook gets run against them.