AI & Creative Tools

What Are We Talking About? AI Is Turning 'Clone the Expensive Software' Into a Weekend Project

Open-source alternatives to commercial creative tools have existed for thirty years. What's new is the timeline: a solo developer with an AI coding agent can now produce a rough clone of seven Adobe applications in a single weekend, instead of the years a GIMP or an Inkscape historically took a volunteer community to build.

Open-source alternatives to expensive commercial creative software are not a new idea. GIMP has been chasing Photoshop since 1996. Blender spent over a decade as a curiosity before becoming a genuine industry tool. Inkscape, Audacity, Krita, LibreOffice — an entire generation of “the free version of X” exists, built slowly, over years, by volunteer communities writing code by hand.

What’s changed in 2026 isn’t the ambition. It’s the timeline. A solo developer can now point an AI coding agent at a commercial application’s feature set and produce a rough, working clone in a weekend — not because reimplementing Photoshop got easier, but because the labor of writing the code got radically faster. This week’s discussion looks at where that history came from, what’s actually happening right now, and the licensing questions the AI-acceleration era has made urgent rather than theoretical.

The old way: clone culture before AI wrote the code

The open-source-alternative tradition has always run on the same logic: expensive, proprietary tools create a gap, and a community fills it slowly, by hand, often over a decade or more. GIMP, started in 1995, is the textbook case — thirty years in, it’s a capable image editor that still isn’t Photoshop, largely because matching a mature commercial tool’s depth, one hand-written feature at a time, is simply slow work. Blender followed a similar arc: open-sourced in 2002 after its original company folded, it spent roughly a decade as a niche tool before genuinely competing with Maya and 3ds Max workflows in professional pipelines.

This site has tracked plenty of that same lineage continuing into 2026, just with AI now doing some of the heavy lifting rather than none of it. ArmorPaint — an open-source Substance Painter alternative — hit its 1.0 release in September after eight years in early access, having rewritten its entire tech stack from Haxe into plain C along the way; its new “Neural Nodes” for photo-to-material generation are a small, locally-run AI feature bolted onto an otherwise traditionally hand-built application. cables.gl, a node-based WebGL visual-programming tool, went fully open source under MIT this year, offering a free alternative to cloud-tethered proprietary visual coding tools. And Halfspace, Matt Keeter’s experimental solid-modeling IDE, carries a pointed detail in its own description: it’s open source under MPLv2, and explicitly “not vibe-coded” — a line that only makes sense as a statement in a year where that distinction suddenly needs making.

Even hardware has its own version of this instinct. Behringer’s SP1200X teases a $200, open-firmware clone of the legendary 1987 E-mu SP-1200 drum machine — not software, but the exact same cultural impulse: an expensive, beloved, out-of-reach tool, recreated and made accessible, with openness as part of the pitch rather than an afterthought.

The new way: a weekend instead of a decade

The clearest example of the compressed timeline arrived this October, when developer Brandon Thomas posted that he’d used AI coding assistance to produce open-source, free clones of seven Adobe applications — Photoshop, Illustrator, Premiere, Lightroom, After Effects, InDesign, and Acrobat Pro — in a single weekend, written in Rust rather than Electron. Thomas was upfront that the result is rough: he called it a “super early alpha,” and outside coverage found the Photoshop clone specifically “missing core features,” with “layer logic issues” and incomplete color management. Thomas’s own prediction — “we’re going to reach 100% feature parity within a month” — was walked back within days on Hacker News to something more measured: “99% parity will take a while, but I’m sure it’ll be measured in months and not years.”

That gap between the initial claim and the walked-back one is, in miniature, the whole story of this moment. An AI coding agent can produce something that looks like seven complete Adobe applications shockingly fast — fast enough that Adobe reportedly noticed the project within days, with no action taken yet — but the last, hardest mile of matching a mature commercial product’s actual depth hasn’t gotten faster at the same rate. Thomas invoked “clean-room replacement” — the legal strategy of building functionality from documentation rather than examining a competitor’s source code — in his public responses, though whether that process was actually followed rigorously is, as one outlet put it, something he would need to be able to prove rather than simply assert.

A similar, slightly further-along story is playing out around Figma. OpenPencil, an MIT-licensed design editor built on the Yoga layout engine with roughly 100 AI tools wired into a bring-your-own-API-key chat panel, can already read and write native .fig files using Figma’s own Kiwi binary codec — genuine interoperability, not just visual similarity. But a reviewer who actually used it for real design work reached a clear verdict: “It probably won’t replace Figma for production team workflows right now,” specifically because it still lacks prototyping and a plugin ecosystem. It’s a tool for solo creators experimenting with an open alternative, not yet a production substitute — which is roughly where most of this wave currently sits: genuinely useful, genuinely fast to produce, and genuinely not finished.

The coding agents making this possible are the same ones making other headlines

This isn’t an isolated phenomenon confined to scrappy weekend projects. The same AI-coding-agent capability is reshaping how commercial software itself gets built. Epic Games reportedly told an internal team — seven months after cutting 20% of staff — to stop writing code by hand entirely and instead edit and guide AI agents for both code and artwork, on an unannounced Disney shooter that’s now delayed indefinitely. NVIDIA’s technical artists are directing Claude and NVIDIA’s own GPT-6 Astra straight into Houdini and OpenUSD pipelines for production 3D work. ControlAnything, a solo research project, has a coding agent write real-time device-integration code on the spot in mixed reality, with no pre-written driver — nine of ten requests reaching a working state in a median of 32.5 seconds.

Inside the creative-tools ecosystem specifically, AI-agent integration has become almost a standard feature rather than a novelty: mcp-for-blender-pro ships an entire manager/worker/reviewer agent hierarchy inside Blender; Embody lets Claude Code assemble TouchDesigner node networks through conversation; ShaderVine was explicitly designed assuming an AI agent, not just a human, might be the one driving the editor. Even Adobe’s own answer to this moment — Firefly AI Assistant — reframes its role as orchestrating Adobe’s existing 60-plus pro tools rather than generating assets directly, a tacit acknowledgment that the “AI agent operating creative software” pattern is now something the incumbents have to answer, not just the clones.

And the pattern cuts both ways on quality: this site covered, back in October, a trade editor’s own account of what AI-generated code is doing to the audio-plugin market specifically — 2026 already among the highest-volume years ever for plugin releases, a market projected to double to $2.51 billion by 2033, and “almost nobody new learned C++” to get there. Volume and velocity are not quality, and the clone wave’s central tension is that AI makes it dramatically easier to produce something that looks like a finished competitor to an established tool, without making it easier to produce something that actually is one.

Openness has always cut both directions, and still does

It’s worth remembering that “open source” and “hobbyist-accessible” aren’t always the same fight, and the current wave inherits an older, thornier set of licensing questions it hasn’t resolved. Ghidra — this site’s hero story from earlier this week — is the sharpest illustration available: a tool the NSA spent roughly two decades building internally, released as open source in 2019, that a community of retro-hacking hobbyists has since turned into the standard toolkit for dumping firmware and running DOOM on hardware that was never meant to run anything but a single cartridge. That’s openness released downward, from an institution with essentially unlimited resources to a community with almost none — a very different shape than a solo developer racing to clone Adobe’s commercial product line in a weekend, even though both get called “open source.”

Meanwhile, the AI-training-data side of openness has become its own contested front. GameMaker’s runtime goes free for commercial use starting January 2027, with source access now included at every pricing tier — a genuine widening of access. But Getty Images, the company that sued over AI training data and then built its own AI generator trained on what it calls “permissioned content,” filed for Chapter 11 the same month, a reminder that “licensed” and “open” increasingly describe entirely different, sometimes opposed, business strategies rather than points on the same spectrum. Smaller AI model releases are threading that needle in real time, too — Gradium’s Phonon text-to-speech model shipped with a deliberate license split between its weights and its training pipeline, an increasingly common pattern for projects trying to be open enough without giving away everything.

Where this actually lands

None of the open-source clones covered here — not the Adobe suite, not OpenPencil, not even mature decade-plus projects like GIMP and Blender in their early years — fully replaced the commercial tool they were built against on day one, or even year one. What AI coding agents have changed is how much of that gap can be closed in public, visibly, before the hard remaining 10% gets any easier. A weekend now produces something that looks, at a glance, like a finished rival; the actual work of matching a mature commercial product’s depth — robust layer logic, full color management, a plugin ecosystem, genuine production reliability — still runs on the old, slower clock that GIMP and Blender’s volunteer communities always worked on. The interesting question for the next year of this isn’t whether AI can produce something that looks like Photoshop in a weekend. It already can. It’s whether that head start actually closes the long tail of parity faster than three decades of hand-written open-source software ever did — or whether it just makes the “super early alpha” stage a lot more crowded.