Library / SDK
starling-build/starling avatar
starling-build/starling

Starling: a Swift Linux desktop that documents its own missing pieces

Starling — a new Linux desktop environment: Swift shell, its own compositor, a Flutter-to-Swift framework port, and first-party apps

417 stars22 forksSwiftApache-2.0

At a glance

What is it?
A desktop environment whose shell, compositor, framework and apps are Swift, running native Wayland clients and X11 apps through its own compositor and X server. Its readme is unusually honest about what does not work: xdg_output is never updated, the portal is incomplete, scaling is pinned, and installing pulseaudio-utils turns a silent Zoom into one that will not start.
Who is it for?
Starling is worth reading and following if you build Linux desktops or graphics stacks, because it names its own gaps instead of leaving you to discover them, and its two most instructive decisions, masking the stock portal and refusing a dependency that fixes one app by breaking another, are the kind of trade-off most project pages hide.
Can I use it commercially?
Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
Is it still maintained?
Yes. The repository received new commits within the last day.
What is it written in?
Mainly Swift, according to GitHub's language statistics.

Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.

From daily snapshots

Star history

  • starling-build/starling417 ★ · +67
Star history
Repository2026-09-192026-11-15
starling-build/starling350417

67 Stars in the last 57 days, measured from the daily GitHub snapshot. Compare with other projects

Editorial analysis

One person, a few months, and a version number that lags

The status section opens with early development at version 0.4.0, and then says the desktop boots as a real session, runs its own compositor and apps, and installs from a .deb on stock Ubuntu 26.04. It also says it is the work of one person and a few months, so expect rough edges, missing pieces and bugs, that nothing here is load-bearing for anyone yet, and that interfaces change without notice.

That last set of sentences is the reason to read this page at all. The next paragraph says the writing is deliberately specific about what is and is not done, because the phrase a desktop environment invites assumptions that would be wrong.

The repository layout makes the shape obvious, with two symlinks standing in for the sibling projects:

text
sdk   -> symlink to a flutter-swift checkout
engine -> symlink to a starling-engine checkout
shell/   the desktop shell: compositor (C Wayland server), window manager,
apps/    first-party apps (Settings, Files, Terminal, ...), one SwiftPM package each

So the Swift port of Flutter's framework arrives as a linked checkout rather than a vendored copy, the engine arrives the same way, the shell is one SwiftPM package, and each first-party app is its own package. bootstrap.sh is what creates those links.

The version number is worth a second look. The status text says 0.4.0 while the newest release tag is v0.5.0, dated 20 September 2026, described as a desktop you can walk into. The readme therefore trails the tags, which is a small thing but a useful signal about how precisely the rest of the document tracks the tree. The two component releases in the same window are the Windows shell at 0.2.0, now on Windows 10 with the first outside reports closed, and the terminal at 0.2.0 with remote workspaces and sessions that outlive the connection.

The session reaches DRM through logind, not through root

The boot path is a login chain, and every step in it is named: gdm3, then gdm-wayland-session, then /usr/libexec/starling-session, then DesktopShellApp --drm.

The security detail is the one to keep. The session is unprivileged: the DRM master and input come from logind through libseat rather than from running as root. On a display server that is the difference between a component asking the session manager for a seat and a component needing elevated privileges to touch the graphics device, and it is why the compositor can run as part of a normal login session at all.

The display side is a conventional DRM and KMS modeset path with GBM and EGL, a hardware cursor, flip-driven frame pacing and multi-output layout. The hardware it was tested on is named as an AMD Radeon 780M and, in a VM, virtio-gpu with virgl.

That hardware list is short and honest. If you are on an Intel or discrete GPU, this page gives you no signal, and the flip-driven pacing claim in particular is the kind of thing that is architecture dependent.

5,700 lines of C against fifteen Wayland extensions

The compositor is C, roughly 5,700 lines, and the protocol list is the real specification. It implements xdg-shell, linux-dmabuf for zero-copy import, viewporter, fractional-scale-v1, pointer-constraints, relative-pointer, text-input-v3, presentation-time, primary-selection, idle-inhibit, cursor-shape-v1, xdg-decoration, xdg-activation, xdg-output and wlr-data-control.

Two entries in that list carry weight beyond the rest. linux-dmabuf is what makes GPU clients zero-copy, and it is the reason a Blender viewport with EEVEE can be driven through the compositor without the frame going through system memory. wlr-data-control is a clipboard protocol that is not in the core Wayland set, so carrying it means the shell is not strictly limited to the standard surface.

On top of the compositor, the window management offers floating and tiling in a master-and-stack arrangement behind one switch, spaces with a Mission Control style overview, drag-move and drag-resize, a dock with running indicators and drag-to-reorder, and a Launchpad.

The framework underneath is a full port of Flutter's Dart framework to Swift with no Dart VM, and it is the part with 137 test files under sdk/Tests, which is the only test coverage figure given anywhere on the page.

xdg_output is announced once and then lies quietly

The first known limitation is precise about the mechanism. xdg_output state is sent once and never updated. A client that asks for an xdg_output gets the logical position and size as they are at that moment; a later mode change, a scale change or an output hotplug does not re-send them, so a long-running client can be left with stale geometry.

The page also names what a fix would require: the resources are not tracked per output, which is exactly what fixing it needs. That sentence is more useful than the bug report, because it tells a reader that this is a design gap rather than a missing refresh call.

The consequence is quiet and it is the bad kind of quiet. Nothing crashes. A window manager client places its popups at coordinates that were correct when the compositor started and are now off screen or on the wrong monitor, and the client has no way to know. Given that there is also no display-mode selection and no way to change scaling, the number of events that can invalidate a stale position is higher than it would be on a finished desktop.

Scaling is effectively pinned to 2.0 for the same family of reasons: fractional values produced blurry text and are not usable yet.

The portal masks the stock one instead of falling back to it

The second limitation explains a decision that would otherwise look like a bug. The portal is incomplete: Settings and FileChooser work, and Inhibit, Camera and Print are absent entirely.

The absence is not a to-do item, it is a mask. The session bus masks the stock xdg-desktop-portal rather than letting it fill those gaps, because the stock daemon has no backend for Starling. If it were allowed to answer, it would serve only its backend-less interfaces while taking org.freedesktop.portal.Desktop away from the shell's own portal, which would break the two that do work.

So the choice is between three portals and two. GTK4 reading org.freedesktop.appearance through the Settings portal is the concrete proof that the working half is real, and the page notes that both working interfaces carry the version property clients probe first.

The ScreenCast portal is the third working one, and it is the interface behind getDisplayMedia and for OBS, which connects it to the screen recording feature rather than leaving it as a protocol name.

Screen recording hands a dma-buf to the hardware encoder

The desktop records itself, either the whole screen or a single window's own content, and writes MP4 into ~/Videos.

The interesting part is the path. Where the machine has a hardware H.264 encoder, the composited frame is handed to it as a dma-buf, so no frame is copied through the CPU, and the cost given is about 0.05 of a core on the tested machine. Without an encoder it falls back to a software path. For a compositor, taking a buffer you already have and giving it to the encoder without a round trip through system memory is the difference between a recorder you can leave on and one you cannot.

The rest is interface honesty: a red dot and a running clock sit in the menu bar for the duration of the recording.

What the limitation list does not say is how the feature behaves across the stale xdg_output problem, which is a reasonable thing to test yourself if window-only recording is what you wanted.

Do not install pulseaudio-utils to fix Zoom

One catalogued app runs with a caveat, and the explanation is the sharpest piece of writing on the page.

Zoom starts but reports no pactl and pacmd found and has no audio. That is deliberate rather than a missing dependency. Zoom segfaults during audio init whenever pactl is on PATH on a PipeWire box with no native PulseAudio daemon, and that PipeWire-only arrangement is the stock Ubuntu 26.04 one. Absent pactl, Zoom skips audio and runs. Present, it crashes, and a working PULSE_SERVER does not save it.

The instruction that follows is emphatic: do not install pulseaudio-utils to fix this, because it trades a silent Zoom for one that will not start.

That is a real maintenance decision with a real cost. Installing the package makes the missing-audio complaint disappear, which is the obvious fix and the wrong one. Other libpulse apps, Chrome, Slack and Teams, are named as working, and the sentence about them is cut off in what is readable here.

A desktop environment that tells you which of your two workarounds will break the app is rarer than it should be.

Ten store entries, generic tiles, and a packaging number

The remaining limitations cluster around polish and reach. There is no screen lock or screensaver. There is no display-mode selection, since the session takes the connector's preferred mode and overriding it means editing the session launcher. The third-party app runtime is opt-in and unshipped: app-run expects a debootstrap'd runtime under /var/lib/starling-apps, and third-party apps are otherwise expected to use their own native packaging.

The App Store catalog holds ten entries, all of which install and launch, but the tiles use category glyphs on purpose, because vendor logos are trademarks and the project ships none of them. Chrome and VS Code show their real icons, read from the host at runtime; the other eight fall back to a generic glyph since only those two ids are wired to that lookup.

The packaging numbers are the ones to plan capacity around: a 50.8 MB .deb that installs on a minimal 26.04 image and pulls 26 dependency packages, with Depends computed from the shipped binaries by dpkg-shlibdeps rather than maintained by hand.

The nine first-party apps are Settings, Files, Terminal with a real PTY, Text Editor, Calculator, App Store, Task Manager, Video Player and Image Viewer, all written against the Swift framework port rather than the original Dart one.

Editorial conclusion

Starling is worth reading and following if you build Linux desktops or graphics stacks, because it names its own gaps instead of leaving you to discover them, and its two most instructive decisions, masking the stock portal and refusing a dependency that fixes one app by breaking another, are the kind of trade-off most project pages hide. It is not something to run as a daily driver on this evidence: one person and a few months of work, interfaces changing without notice, and no screen lock. Try it in a VM with virtio-gpu, which is what it was tested on, and read the known limitations before you trust it with a display you care about.

Frequently asked questions

What is Starling and how does its session start?

Starling is a Linux desktop environment written in Swift, with its own Wayland compositor and its own X11 server. The session boots through gdm3, gdm-wayland-session, /usr/libexec/starling-session and DesktopShellApp --drm, unprivileged, with DRM master and input coming from logind via libseat.

Which Wayland protocols does the Starling compositor implement?

About 5,700 lines of C covering xdg-shell, linux-dmabuf for zero-copy import, viewporter, fractional-scale-v1, pointer-constraints, relative-pointer, text-input-v3, presentation-time, primary-selection, idle-inhibit, cursor-shape-v1, xdg-decoration, xdg-activation, xdg-output and wlr-data-control.

What are the known limitations of Starling?

xdg_output state is sent once and never updated, so clients can be left with stale geometry. The portal has no Inhibit, Camera or Print. There is no screen lock or screensaver, scaling is effectively pinned to 2.0, and there is no display-mode selection without editing the session launcher.

Why does Zoom have no audio in Starling?

Zoom segfaults during audio init when pactl is on PATH on a PipeWire machine with no native PulseAudio daemon, which is the stock Ubuntu 26.04 arrangement. Without pactl it skips audio and runs, so do not install pulseaudio-utils to fix it: that turns a silent Zoom into one that crashes.

Official sources

  1. License: Apache-2.0
  2. Project website
  3. README
  4. Releases
  5. starling-build/starling on GitHub
Add this badge to your README

If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.

Add this badge to your README

markdown
[![Hysen Labs](https://hysenlabs.com/badge/starling-build-starling.svg)](https://hysenlabs.com/projects/starling-build-starling)

Star history chart for your README

markdown
[![Star history](https://hysenlabs.com/chart/starling-build-starling.svg)](https://hysenlabs.com/projects/starling-build-starling)

Was this useful?