The AI Spotlight Follow
He Cloned Photoshop With AI in 24 Hours. Now He Says 4,000 Developers Are Helping Him Clone Everything Else
AI Coding

He Cloned Photoshop With AI in 24 Hours. Now He Says 4,000 Developers Are Helping Him Clone Everything Else

After his AI-built Adobe and Office clones went viral, ArtCraft's Brandon Thomas laid out the bigger plan: rebuild the closed software world as open source, in Rust, run by coding agents. Here's how the 24-hour Photoshop happened, what came next, the rules he uses for agents versus humans, and the legal and quality questions he can't wave away.

Advertisement

A few weeks ago, a developer in Atlanta posted some open-source apps that looked almost exactly like Photoshop, Premiere and InDesign. Free. Built largely by AI. Rewritten in Rust.

He hadn't even posted them to Hacker News yet. Someone found them on Reddit. Two days later, by his account, The New York Times was asking for an interview.

The developer is Brandon Thomas, and his company is ArtCraft. We've already covered the second act, when he did the same thing to Microsoft Word, Excel and PowerPoint. In a recent talk, thrown together, he admitted, late the night before after a sleepless month of VC meetings, he laid out the much bigger plan behind it.

The plan, in his words: "open source the entire world."

Advertisement

It sounds like a joke. Here's why a lot of people are taking it seriously, and why others think it's legally and technically shakier than it looks.

Who's behind it

Thomas has been an engineer for about 15 years.

  • Eight years at Square, working on heavy infrastructure: online payments, the authentication stack, "anything that could take down the company." Eight years of carrying the pager.
  • Seven years in startups, which he describes as a different world of risk, learning what the market wants and correcting your own assumptions.
  • A long side life in film and art in Atlanta, a big film state: making films out of college and building laser projectors that played video games on the sides of skyscrapers.

That mix explains ArtCraft. Its customers have always been creatives: artists and filmmakers.

The argument: we lost the open web

Thomas's case starts with nostalgia, but it's a sharp argument.

In the 1990s and early 2000s, he says, the internet was built for tinkerers:

  • Chat was open. IRC was like Discord, except you could run your own servers and automate them however you liked. Commercial messengers didn't care if you used third-party clients.
  • Data was open. Websites had exports and APIs. People built "mashups" of Reddit, Flickr and Twitter that trended on Slashdot before Hacker News existed.
  • You owned things. You could buy an MP3 and feel it was yours, not license it.
  • Creativity was everywhere. Flash animators, sprawling forum communities on phpBB and Invision, the early ethos that produced Wikipedia, and MySpace pages where you could rewrite the HTML and CSS.

Then came Web 2.0, broadband and smartphones. That brought everyone online, which was good: our parents and grandparents needed usable products. But it also "rounded the edges." Protocols were replaced by platforms. APIs disappeared. App stores became gatekeepers, and we never got a phone ecosystem where you could distribute freely the way you could on a desktop.

And as companies won market share, they discovered pricing power. Subscribing to every streamer, Thomas says, turned Netflix back into cable.

The cancellation fee that started it

The personal trigger was a dark pattern many creatives know.

Adobe's annual plans billed monthly carry an early termination fee if you cancel partway through, which Thomas describes as buried in the fine print and costing up to about $200. For someone who uses Premiere for a weekend once a year to make a short film for a festival, that can turn a small project into a surprise bill.

Ars Technica reported that Thomas said he was inspired to build the clones after being bitten by that cancellation fee "one too many times." Adobe's cancellation terms have also drawn regulatory fire: in 2024, the US Justice Department sued Adobe on behalf of the FTC over its early termination fees and cancellation process. Adobe has since agreed to a $150 million settlement, according to the Justice Department, including requirements to disclose fees upfront and offer simple cancellation.

To Thomas, it's a symbol of a web that went from open and free to "extractive" and "rent-seeking." His word for it is a popular one: enshittification.

The surprise: artists started coding

ArtCraft didn't begin as an Adobe killer. It started as a professional-grade AI creative tool for filmmakers, a desktop app where generated files lived right next to the editor, aimed at pros rather than casual users.

And it was fully open source, apps and infrastructure. Thomas says his advisers thought he was crazy: artists don't know or care what open source is.

Then the team saw something nobody expected. Artists were vibe-coding their own workflows on top of ArtCraft's software.

His example: micro-drama studios doing fulfillment for top micro-drama startups, $30,000 to $50,000 of work a month, with big teams. No product fit their exact delivery needs, and nobody was going to build one just for them. But they had in-house people who knew Claude Code. They took the APIs and generated the interface they needed.

That raised the question that led to everything else: what if the apps creatives already rely on, like Premiere and Photoshop, were open too? Teams could strip out everything they don't use, constrain each app to a specific workflow, and do it five or ten times, once for each artist on the team.

Photoshop in 24 hours

So they tried. Here's the process Thomas described:

  1. Lean on what the model already knows. Frontier models have read countless Photoshop tutorials. The team's prompt essentially said: this is the basic shape of a photo editor; you know what Photoshop is.
  2. Write a huge, structured prompt. Separate the UI and UX from the core. Make everything MCP- and command-line-compatible, so AI agents can drive it.
  3. Plan every tool. Specify how each tool works, the data structures and the file-format importers.

Within 24 hours, he says, they had "a pretty good approximation" of Photoshop. It was functional, it felt familiar, and it started 10 times faster than the original.

It's not finished. By Thomas's own measure, the clones were at about 50% of Photoshop's feature surface and had just passed 60%. He invokes the Pareto principle: the remaining percentage points are the hard ones, but he believes 99.99% parity is reachable eventually. Ars Technica noted that early on he promised 100% parity "within a month," then softened it to months rather than years.

For now, the team is deliberately copying Adobe's look and feel so artists can keep their muscle memory. New, more agent-native interfaces will come later. Hitting parity is also meant to prove something: that a real, complex product can be built almost entirely by agents.

Ars reported the initial suite covered seven apps mirroring Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign and Acrobat Pro, built with Anthropic's Claude Opus 5.5, under MIT/Apache licenses, with funding coming from paid access to ArtCraft's own AI image and video tools.

The viral month

What happened next, by Thomas's telling:

  • ArtCraft's Discord grew from about 3,000 members to more people than the Discords of several well-known AI creative startups.
  • He says ArtCraft's repositories have passed 300,000 GitHub stars combined, with about 4,000 contributors. The PhotoCraft repo alone gained roughly 77,000 stars in about a month, according to a GitHub tracker.
  • He claims around 1.7 million downloads a day, and accelerating.
  • A quarter of users are on Linux, a platform he says has been badly underserved by professional creative and office apps.

Treat those figures as his claims; stars and downloads can be inflated and don't equal active users. But the attention is undeniable.

The most surprising reaction came from anti-AI artists, people who had spent years pushing back against generative AI. Many, he says, felt mistreated by Adobe and other tech giants, and they've started using the apps, filing bug reports and feature requests. Thomas sees a chance to teach them what MCP is and how to create with agents.

Beyond Adobe

ArtCraft and its community haven't stopped at Creative Cloud. According to Thomas, the projects now include:

  • CAD tools, including a community project called SolidCraft, led by engineers with an open engineering kernel and MCP support for fast 3D-printing projects.
  • A Flash-style animation tool, and support for Adobe Animate-style workflows after animation studios asked them to pick it up as Adobe's own support wound down.
  • Several digital audio workstations.
  • Core Microsoft Office apps and a Google Docs alternative.
  • A Figma alternative, plus a server component for team collaboration using consensus algorithms like Raft and Paxos for multi-user editing.
  • A PDF app he calls probably the best of the lot, fast and smooth, running on Linux and even NetBSD.

The community is building more than 300 apps, he says. ArtCraft is breaking its UI into reusable 2D and 3D component libraries and building a tool so non-programmers can contribute. There's a desktop launcher to manage everything across operating systems. Apps have been localized into languages the originals didn't officially support; he cites Arabic in an InDesign-style app.

And everything, including the infrastructure and data model, is open source. Nothing is kept proprietary.

Apps that agents can drive

One design choice runs through everything ArtCraft builds, and it may matter more than the cloning itself: every app is MCP- and CLI-first.

MCP, the Model Context Protocol, is the open standard that lets AI agents discover and use a tool's functions. A CLI is a command-line interface. Together, they mean an AI agent can operate the app directly, not by clicking through menus like a human, but by calling its functions.

Thomas says you can command these apps and watch them build and assemble creative projects in front of you. Imagine telling an agent to cut a trailer from a folder of clips, color-grade it, add titles and export versions for three platforms, inside an editor built from day one to take orders from software.

That flips the usual relationship. Most professional apps were built for hands and mice, with automation bolted on later. These are built for agents first, with a familiar interface on top for humans.

Why Rust

Thomas calls Rust "probably essential" to the project's success. His reasons:

  • Native performance. Others have built clones in Electron, but high-quality creative apps need native speed.
  • Reliability. He points to Google's published findings on Rust's low defect rate.
  • Fast startup.
  • Portability. Every major OS, tablets, NetBSD, and the web through WebAssembly.

Do you need to know Rust to start? No, he says. Many on the team had never used it. "Your agents will drive it just fine." The real pain is compile times. And yes, Rust can do rich GUIs now.

He also listed the unglamorous problems of running a viral, agent-heavy project: permissions, GitHub limits, scaling, and CI falling over under the load. The team is moving to its own CI on colocated servers where RAM is cheaper.

The rule: when to trust agents, when to use humans

This was the most practical part of the talk.

Thomas's rule of thumb:

  • Use humans for back-end work and anything involving schema migrations or API design, decisions that are hard to undo and ripple through everything.
  • Unleash agents on desktop apps, games and anything essentially self-contained. Give them full confidence. You don't need a human in the loop at every step; direct where things go upstream and let them run.

His most eye-popping example: someone in the community has 40 machines rebuilding Unreal Engine in Rust.

Autocraft: the software factory

The final reveal was Autocraft, which he describes as a fully autonomous software factory that ArtCraft plans to open-source. As described, it:

  1. Ingests issues from the community.
  2. Triages them.
  3. Does agentic development to fix them, with open-source AI code review in the loop.
  4. Keeps engineers on task.
  5. Records every decision as training data for evaluations and post-training.

That last point is the quiet ambition. A factory that learns from its own decision history gets better at running itself.

He closed with the mission: rebuild all software and platforms, turn them back into proper protocols, and own our technology again.

Why this matters

Even if you're skeptical of every number, three things are real.

Cloning is cheap now. What once took a big team years is being approximated by a tiny team with agents in days. That changes the economics of every software category.

Incumbents' moats are shifting. Features are easy to copy. What's harder: file-format fidelity, reliability, ecosystems, support and trust. Expect the big vendors to compete there.

Open source gets a new engine. If agents handle much of the grunt work, small communities can maintain huge codebases, and platforms like Linux get apps they never had.

The serious objections

  • Is it really "clean room"? Ars Technica readers questioned it. Traditional clean-room reverse engineering uses separate teams for specs and implementation. One AI-driven pipeline doing both, critics argue, isn't the same.
  • Trade dress. Copying Adobe's look and feel is a deliberate choice, and Ars noted it could invite legal trouble even where the code is original.
  • Quality and parity. "60%" of Photoshop is a long way from what professionals rely on daily, and early users reported plenty of shortcomings.
  • Sustainability. Viral projects often fade. ArtCraft says paid AI services fund development, but maintaining hundreds of apps is enormous.
  • Indie collateral damage. Some argue cheap AI cloning will hurt small software makers first, since they can't outrun a one-person clone shop.
  • Security. Agent-written code at this volume needs serious review, especially for apps handling files and collaboration servers.

Practical takeaways

  • Try before you switch. Test the PDF or photo apps on your real files; check import and export fidelity.
  • Keep the originals for client work until formats round-trip reliably.
  • Steal the workflow. Huge structured prompts, MCP/CLI-first design and Rust are a template for any agent-built app.
  • Copy the rule. Humans for schemas and APIs; agents for self-contained apps.
  • If you sell software, assume your feature list can be cloned. Invest in what can't be.

FAQ

What is ArtCraft?

A company and open-source community led by developer Brandon Thomas that started as an AI creative tool for filmmakers and now builds free, Rust-based alternatives to Adobe, Microsoft and other closed apps, largely with AI agents.

Did he really clone Photoshop in 24 hours?

He says the team had a functional, familiar approximation within 24 hours. By his own measure, it now covers about 60% of Photoshop's features, not all of them.

Reverse engineering is generally allowed, but critics question whether an AI pipeline qualifies as clean room, and closely copying an interface could raise trade dress issues.

Why Rust?

Native performance, reliability, fast startup and portability, including the web via WebAssembly. Agents can write it even if you can't.

What is Autocraft?

An autonomous software factory ArtCraft says it will open-source: it ingests and triages issues, develops fixes with agents and uses its decisions as training data.

The bottom line

ArtCraft's grand promise, open-sourcing the entire world, will run into lawyers, file formats and the long tail of features that make professional software professional.

But the core demonstration has already landed. A small team with AI agents can now produce credible versions of software that once took decades and thousands of engineers. Whether ArtCraft wins or not, every software company now has to answer the same uncomfortable question: if your app can be cloned in a day, what exactly are customers paying for?

Sources

The Creator of an AI Hacking Tool Shut It Down. At Least 1,465 Servers Are Still Running ItUp next · 12 min readThe Creator of an AI Hacking Tool Shut It Down. At Least 1,465 Servers Are Still Running ItARTEX was built to help companies test their own defenses. Then it was used against South Korean banks, a Chinese state-linked group turned it on Microsoft 365 customers, and its developer walked away. Copies kept spreading anyway. Here is what the logs show, how much the AI really did, and what actually stopped it.Keep reading →