Some Linux desktop environments were developed over decades, others in just a few years. Starling, which made the rounds online this week, claims to have been built by just one person in six months using AI coding tools.
What’s more, a preview build is available as a .deb to install on Ubuntu 26.04 LTS. Yup, this isn’t a ChatGPT mockup but a real working demo.
According to the project’s GitHub and website1, a developer used AI2 to rewrite Flutter’s framework layer from Dart to Swift, and left its C++ rendering engine unchanged. A traditional desktop uses a windowing system but here the engine drives the display via DRM/KMS.
A Wayland compositor and X server, built for this project, take each app’s window and hand it to the framework directly off the GPU, with no copying involved. The framework then treats that window the same way it treats its own dock or buttons or window bars.
Now, zero-copy GPU buffer sharing is a Wayland protocol, and mixing native and X11 clients in what Xwayland does. What’s actually new here is narrower: treating that composited ‘buffer’ as a widget inside the framework’s own layout system.
Novel? As a feat, perhaps. But enough to hang an entire desktop’s USP on?
Why does Starling exist?
I’ll be honest: I’ve been writing and rewriting this post for several days because it is hard to get a handle on what Starling is aiming to be and who it’s for because any human impetus is buried beneath a mountain of ‘profundity slop’.
The Starling website and GitHub pages are full of breathlessly hollow statements and narrative framing, and the unrelenting rhetorical glibness comes at a headache inducing rate.
One quote on why this effort is so supposedly remarkable jumped out at me:
A curious person with a bold idea is not new. Being able to act on it alone is.
Starling.build website
This is brilliant example of what I called “polyester prose”. Words LLMs arrange to look flashy, but whose logic is frayed at the edges (and leaves me feeling itchy). I wonder if the AI was prompted to write a manifesto in the style of those tai-chi-walking scam ads, but with less sincerity.
To wit, the Enlightenment desktop exists. Its founder built the first version alone in 1997, and it grew into a full desktop running on its own toolkit – decades before AI tools existed.
Starling’s own “why” page even briefly mentions this. But instead of zoning in to explain why or how its approach differs, it ignores its existence and reverts to hyping itself as the result of an AI-gifted super-power that finally lets people act on their bold ideas alone.
Parodic TedTalk-isms carry over into this dozy: “if a single person can now reach this point, then the constraint on Linux desktop development was never imagination and never taste. It was labour. And labour is exactly what just got cheap”.
How about commitment?
A bold idea or a repetitive one?
I’d charitably describe Starling’s look as “Temu macOS Mojave in dark mode”
Screenshots of Starling on the website show a modern, if traditional, desktop layout with a UI style I’d charitably describe as “Temu macOS Mojave in dark mode”.
Dock, panel, clock, window switcher. There’s little user-facing that couldn’t be recreated on any Linux desktop install using a few apps and a stale macOS GTK theme.
What benefit there apparently comes from the fact tiling, spaces and the dock glass-effect work on apps like Google Chrome without the need to explicitly write code to support them. A suite of first-party apps are available, including a file manager, terminal and calculator.
And that’s it, really.
Once you’re past the vibe-coded engineering newspeg, it’s just a patchwork of imitation atop a mountain of overwrought value statements over a code approach that isn’t “bold”, much less any other superlative.
But… Could it take off?
Starling looks nice at first glance. It’s doing something novel on the backend, fine. My cynicism – the provenance of and ethics of its AI generated code notwithstanding – is more on the fleeting nature vibe-coded projects have.
It could go on to become a properly maintained Linux desktop that attracts users and contributors en masse, and mature in new and interesting way. Or, like a lot of “AI slop”, it may just exist because it can and will be promptly abandoned as soon as the interest wears off.
AI coding nixes constraints, and constraints do matter; friction is the 9/10ths of the product, as I heard someone say on a podcast recently.
The fundamental reason why some software is good and some software is not, is that someone, somewhere had to make a choice because of a constraint (self-imposed or otherwise) and carry the ‘cost’ of making that choices as the maintainer.
‘What costs nothing to make, costs nothing to abandon’, to mangle an aphorism that rings true for vibe-coded software.
You can try the Starling desktop on Ubuntu 26.04
If you want to try Starling yourself, you can. Download the .deb from the project GitHub and install it on Ubuntu 26.04 (not a container; Starling needs a real AMD GPU to work). Then, pick the Starling session at GDM login screen and log in as normal.
If you’re a DE-hopper, look at the list of known issues on the GitHub as it lacks support for Intel and NVIDIA GPUs, third-party apps, working audio, a screen lock.
I will add a note of caution.
Installing random deb packages with root permissions on a stable system, originating from developers you don’t know or trust is a risk. If software is a double-whammy of pre-stable state and vibe-coded, 2x the risk. One person who tried it filed a bug to say it corrupted their install.
- A truly bold looking website? Of course not, it uses the same generic AI design aesthetic that now blankets much of the modern web. So bored of the generic look, I’m beginning to reappraise the visual skill from the Geocities era of web design. ↩︎
- Amusingly, or depressingly depending on your POV, the developer claims sole credit for the idea, saying “the bet was mine, the judgement was mine”, etc but reaches for “we” when recapping the implementation, so it’s “we took the engine’s C core”, “we ran it on…” etc. Perhaps I should tell my microwave “we” cooked a great meal together tonight using “my” recipe. ↩︎