Five Adobe-Scale Apps, Five Weeks, Pure Rust

Something specific happened on GitHub in the last five weeks, and it is easiest to see as a list of star counts.
storytold/photocraft, an open-source photo editor written from scratch in pure Rust: 76,436 stars since September 30. storytold/wordcraft, a word processor with a ribbon, styles, tables, track changes, references and mail merge, reading and writing .docx: 17,139. storytold/solidcraft, parametric 3D CAD: 13,295. storytold/cadcraft, CAD drafting in the AutoCAD sense: 7,521. storytold/gridcraft, a spreadsheet: 7,074.
Five applications that took Microsoft and Autodesk decades, rebuilt as clean-room Rust implementations, in a bit over a month, by what appears to be one small team. And a sixth repo, replica-skill, with 7,254 stars, offering "eleven free Claude skills that clone any app: reverse-engineer it, rebuild it."
The interesting question is not whether an agent can write a spreadsheet. It obviously can. The interesting question is what actually got cheaper, and what did not.
What "clean-room" is doing in that sentence
Clean-room implementation is a legal term of art. It means you build a compatible product without copying the original's source, usually by having one team write a specification and a second team implement from that specification alone. It is how Compaq made the first IBM-compatible PC, and it is why "compatible" software is legal at all.
The reason it matters here is that it explains the shape of these repos. They are not forks. They are not decompilations. They are reimplementations against a behavioral spec, which is exactly the kind of task a language model is good at, because the spec is text and the implementation is text and the test is "does the output file open in the real thing."
photocraft's README is explicit that it is a clean-room reimplementation, MIT/Apache-2.0 dual licensed. wordcraft's says the same, and adds the detail that actually matters for adoption: it reads and writes .docx, runs natively on macOS, Windows, Linux and BSD, and runs in the browser via WebAssembly. The compatibility surface is the product.

The corpus is the tell
Buried in the org's repos is photocraft-corpus, 162 stars, described as "test data for photocraft."
That repo is the most important thing in this story and the easiest to skip past. Reimplementing Photoshop is not hard because the UI is hard. It is hard because the file format is a museum of thirty years of accumulated decisions, and because "correct" is defined by whether a specific document renders the same way it does in the original.
You cannot get there with a prompt. You get there with a corpus: a pile of real files, a way to render both implementations, and a diff. Once you have that, the task decomposes into thousands of small, checkable, agent-sized problems. Without it, you have a demo that looks right until someone opens a file with a layer mask and a smart object.
This is the pattern I keep seeing in agent-built software that actually ships. The scarce input is not code generation. It is a specification precise enough to check, and a test harness that can tell you "no" without a human in the loop.
MCP as the distribution channel
wordcraft's README carries a badge that would have been strange two years ago: "Agent-drivable over MCP / CLI."
Read that as a product decision. These applications are being built for two kinds of users simultaneously: a person who wants a word processor, and an agent that needs to produce a document. The second user does not care about the ribbon. It cares that insert_table exists and returns a structured result.
This is a genuinely new requirement, and it explains why a from-scratch implementation can compete with a mature one. Microsoft Word is not going to expose a clean MCP surface for arbitrary agents, because its business model depends on the document living inside its ecosystem. A 17,000-star Rust word processor with an MCP server has no such conflict. It can be the thing an agent writes into.
The same logic applies to gridcraft and the CAD tools. The value is not "free Excel." The value is "a spreadsheet an agent can drive deterministically, on a machine you control, without a license server."
What got cheaper, precisely
Three things, and it is worth separating them because they have different implications.
Boilerplate got free. Menu trees, preference panes, keyboard shortcut tables, file dialogs, undo stacks. This is 60% of a desktop application's line count and roughly 0% of its difficulty. Agents are extremely good at it. Five years ago this was a month of work per app.
Format compatibility got tractable. Not free — tractable. With a corpus and an agent that can iterate against a failing test, the long tail of .docx edge cases becomes a grind rather than a research project.
Cross-platform got cheap in a specific way. Rust plus WebAssembly means one codebase produces a native binary and a browser build. That combination used to require two teams.
What did not get cheaper: taste, performance engineering, and the political work of deciding what "compatible" means. A Rust photo editor that opens PSD files correctly but feels like a database client will not displace anything. The 76,000 stars are a bet that it will not feel like that, and the bet is unproven.
The uncomfortable questions
Is this legal? Clean-room reimplementation of an interface is generally lawful, and the licenses here are permissive. But "generally lawful" is doing a lot of work, and the history of software clones is a history of lawsuits that turned on details nobody predicted. The .docx format is an ECMA standard, which helps. PSD is not.
Is this sustainable? A 76,000-star repo created five weeks ago has no maintainer bench, no release cadence, and no answer to "who fixes the bug in the smart-object path in 2028." The star count measures interest, not durability. Every one of these projects will be judged in two years by whether it still opens the files it opened in October.
Does it matter that a model wrote it? Less than people think, and more than they want. Less, because the output is source code and source code is judged by behavior. More, because the economics of "one team, five applications, five weeks" only works if the maintenance cost is also transformed, and there is no evidence yet that it is.
The thing to watch
Not the star counts. Watch the corpus repos, the test harnesses, and the issue trackers six months from now.
If photocraft-corpus keeps growing and the issue count stays flat, then something real has changed about how complex software gets built, and the five-week timeline was not a stunt. If the corpus stalls and the issues pile up, then what we watched was a very impressive demonstration of code generation against a specification that was easy for the first 80% and impossible for the last 20% — which is the same story software has always told, just faster.
Either way, the honest summary of this month is not "AI replaced Adobe." It is that the cost of building a credible clone of a mature desktop application has dropped enough that people are doing it for free, in public, and the constraint has moved from typing to specifying.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.