know.2nth.ai Tech The malleable workstation
tech · Platform & desktop

Everything above the OS became editable. Now the OS is too.

On 14 August 2026 DHH shipped Omarchy 4 “Quattro” — a Linux desktop that treats coding agents as system citizens rather than programs you happen to run. The distro is the visible part. The argument underneath it is the interesting part: a machine made entirely of plain text files is the only kind an agent can genuinely rebuild, and that rules out macOS and Windows by construction. This is what that claim gets right, what it costs you, and how it stacks against the two other serious answers to the same problem.

Arch + Hyprland Omarchy 4 “Quattro” Quickshell MIT Rolling release

Agents stopped being apps and became system citizens.

Omarchy 4.0.0 landed on 14 August 2026 and is the largest change since the project started. The desktop shell — bar, launcher, menus, notifications, lock screen, polkit agent — was rebuilt as a single Quickshell process, replacing Waybar, Walker, Mako, SwayOSD and hyprlock. That is the engineering headline. The positioning headline is what got wired in alongside it.

Agent, chosen once

A default coding agent

Pick one — Claude Code, Codex, Gemini, Copilot, OpenCode, Crush, Grok or Pi — and launch it from anywhere with SUPER + SHIFT + CTRL + A or the terminal alias a.

Bar widget

Model spend, next to the battery

A widget tracking how much of your Claude or Codex allowance you have burned. Sitting in the same row as the clock and the battery indicator.

Crash path

Core dumps go to the agent

When a process segfaults, the OS offers to brief your agent with the core dump instead of leaving you a file in /var/lib/systemd/coredump.

Plumbing

Lua configs, smaller ISO

Hyprland config converted to Lua for 0.56 compatibility. ISO under 6 GB (down more than a gigabyte), installs ~30% faster — sub-minute is achievable.

Also in the release: the launcher and menu merged into one surface on SUPER + SPACE, notification history replay, autogenerated theming expanded from 8 base colours to 24, a community plugin architecture for custom widgets and bars, three new themes, and a small suite of first-party apps (Omawrite, Omacut, Omacalc).

None of the agent features are technically hard. A usage widget is an API call. Piping a core dump into a prompt is a shell script. The point is not difficulty — it is that they are defaults. Putting agent spend next to the battery indicator is a statement about what the operating system now thinks it is for, and no other desktop OS has made it.

Open source promised malleability. Almost nobody collected.

The distro follows from an essay. In “The malleable computer” (15 April 2026) DHH makes a two-step argument, and it is worth separating the halves because one is much stronger than the other.

Step one — hard to argue with
  • Open source always granted the right to change the code
  • The barrier was never the licence — it was effort
  • Agents collapse the effort of a small, local change
  • So the right to fork finally has cash value
Step two — the load-bearing claim
  • × On macOS, the shell, WM and notifications are Apple's
  • × On Windows, same story with a different owner
  • × You get surface hacks, not system changes
  • × Therefore real malleability only exists on Linux

His own framing of step one: open source promised users could change their code, but “hardly any of them ever did — it was simply too hard.” Step two is the sharper edge: “you can only do this on Linux.” He pushed the same argument at contribution gatekeeping two months later in “Let the agents democratize open source” (1 June 2026).

Where it holds. If your objection to Linux on the desktop was ever “I do not want to spend a weekend configuring it,” that objection is materially weaker in 2026 than it was in 2016. The configuration surface has not shrunk — the cost of navigating it has.

Where it has a hole

Malleability you cannot reason about is a liability, not freedom. A machine an agent has quietly rewritten two hundred times is a machine nobody can debug — unless every change is versioned, reviewable and reversible.

That is a configuration-management problem, not an operating-system problem, and Omarchy does not solve it for you. Its answer is “it is all dotfiles, so put them in Git.” That is a genuinely good answer. It is also entirely on you, from day one, and it is exactly the gap the two alternatives in section 04 were built to close.

Arch, Hyprland, and a very strong set of opinions.

Strip the marketing and the technical description is short. Omarchy is not a fork of Arch and it invents no new package format. It is stock Arch Linux, the Hyprland tiling Wayland compositor, the Quickshell desktop shell, a curated package set, and a large, carefully-made pile of dotfiles — delivered as an ISO so you never assemble it yourself.

The shape of it

  • Base: Arch Linux, rolling release. Upstream lands on your machine the day it lands, breakage included.
  • Compositor: Hyprland (Wayland, tiling, keyboard-driven), configured in Lua since v4.
  • Shell: Quickshell — one process for bar, launcher, notifications, lock screen.
  • Licence & source: MIT, github.com/omacom/omarchy, ~35k stars.
  • Lineage: shipped June 2025 as the Arch sibling to Omakub, DHH's Ubuntu setup. Not a replacement for it.
  • What it is not: not immutable, not declarative, not atomic, not multi-user, not a server OS.

The standard criticism is that this is Arch plus Hyprland plus dotfiles, and the innovation is presentation rather than technology. That criticism is factually correct and mostly beside the point. Presentation was the missing piece. Every part of Omarchy existed and was free for years; what did not exist was someone with taste making all of the choices and shipping the result as one install. A hundred thousand people downloaded it in the week after v4 — not because the technology is new, but because the decisions were already made.

Taste is a real contribution. It is also the thing that makes the project fragile in a specific way, which is section 05.

Three ways to build a machine an agent can rebuild.

“Agent-editable OS” is not a single design. There are three mature answers in 2026, and they disagree about the most important question: what happens when the change is wrong. Omarchy optimises for the edit being easy. NixOS optimises for the edit being reproducible. Immutable images optimise for the edit being undoable.

OmarchyNixOSImmutable images
ModelImperative dotfiles + pacmanDeclarative — whole system from one expressionOS shipped as a container image (Bluefin, Aurora, Silverblue)
Reproducible?No. Same install script, drifting resultYes, by designYes, at the image layer
UndoWhatever snapshotting you set up yourselfBoot the previous generationRoll back to the previous deployment
Agent ergonomicsBest. Plain text the model already understandsStrong in theory, one file to edit — weaker in practice, models invent Nix optionsConstrained on purpose. The agent cannot casually change the base
How it breaksUpstream rolls, something stops workingThe config fails to build — it fails safeRarely. Layering extra software fights you instead
Learning costLowHigh — a language, not a config fileLow until you need something off-image
SuitsA solo developer workstationTeams that need identical machinesFleets and security-sensitive estates

The honest summary: Omarchy has the best agent ergonomics of the three and the weakest safety net. Those are the same property viewed from two sides — nothing stops the agent from editing anything, which is the whole appeal and the whole risk. If you run it, the dotfiles go into Git on day one, not the first time something breaks. Do that and it holds. Skip it and NixOS is the more intellectually honest version of the same idea, while an immutable image is the safer one.

The gotchas the landing page does not lead with.

None of these are disqualifying on their own. All of them are things you would rather learn before the install than during it.

1 · It wants the whole disk, and the firmware relaxed

The documented install shape is a dedicated drive with full-disk encryption. The v3 manual was blunt about the prerequisite: “You must turn off Secure Boot and/or TPM in the BIOS.” The v4 manual has improved on this — there are now dedicated chapters for dual-boot and unattended installs — but the mental model is still one machine, one owner, one disk. In an estate where Secure Boot is a compliance control, this is where the conversation ends.

2 · It assumes an expensive screen

Defaults target 2x retina-class displays — 218 PPI and up. On a standard 4K panel or a 1080p laptop you are adjusting scaling on day one. Ironic for a project whose economic pitch is that you do not need premium hardware.

3 · Rolling release means upstream can ruin your Tuesday

This is not hypothetical. After Hyprland 0.53 landed in January 2026, users hit config errors, broken themes, cursor sync problems and crashes that stopped the screen loading; 3.4.0 shipped to clean it up. The distro is downstream of a fast-moving compositor, and that is a structural property, not a bug that got fixed.

4 · Single-user by design, and no fleet story

No MDM, no policy layer, no multi-user or server posture. Full-disk encryption on by default is a genuine POPIA-friendly starting point, but “encrypted” is not “managed.” If a client needs to demonstrate control of the endpoint, this does not do it.

5 · The governance got concentrated, fast

The Omacom Foundation launched on 21 August 2026 with $10 million, Drew Houston and Peter Steinberger as founding patrons, and immediately became the exclusive sponsor of Hyprland, plus a funder of Quickshell and mise. Read it both ways. The upside is that the compositor and shell your desktop depends on are now funded rather than volunteer-burnout-limited. The risk is that one organisation now ships the distro and funds its two most important upstreams — Omarchy's opinions become the ecosystem's defaults with no friction in between.

6 · The non-technical objections are real objections

It ships some closed-source applications, which sits awkwardly against the freedom argument. And DHH is a polarising figure — in some organisations that is a procurement conversation rather than a technical one. You do not have to agree it should matter to plan for the fact that it sometimes does.

Where it earns its place, and where it does not.

The split is cleaner than most: it is about who owns the machine and what happens when it breaks on a Wednesday morning.

Run it when

  • It is your own development workstation and nobody else has to support it
  • You already live in a terminal and a tiling WM is an upgrade, not a tax
  • You want agents editing the machine — and you will keep the dotfiles in Git from day one
  • You are buying x86 hardware and would rather skip the Apple premium
  • Single-purpose builds: demo rigs, workshop machines, a dedicated agent box
  • You want the Linux desktop question answered by someone with taste instead of answering it yourself

The argument that lands here is the hardware bill, not the ideology.

Most of the Omarchy discourse is about freedom. Locally, the part that actually changes a decision is what a competent developer machine costs in rands — and what it does when the power is out.

The hardware maths is the pitch. The project has driven a visible move toward cheap x86 mini PCs and mid-range laptops. DHH's own comparison ran the HEY test suite — a real workload, not a synthetic benchmark — on a roughly $300 mini PC against an M4 Pro Mac and found the gap surprisingly narrow. Convert that spread into ZAR with import duty and the Apple premium on top, and the same argument that is interesting in Copenhagen is material here. For a small team kitting out four or five developers, it is the difference between one machine and two.

Battery matters more here. He clocked a Panther Lake Dell XPS 14 running Omarchy at about 1.4 watts idle draw — 47+ hours of standby and roughly 16 hours of mixed use. In a market where load-shedding is a planning assumption rather than an anecdote, a laptop that survives a working day off a UPS-free desk is a real operational feature.

Plan a second machine, not a partition. If your client work runs through Windows and M365 — and for most SA corporate delivery it does — do not try to make one laptop serve both worlds off one disk. Buy a separate box. Given the price point above, that is a far more reasonable sentence in 2026 than it used to be.

Watch the bandwidth. A 6 GB ISO plus a rolling release that updates continuously is fine on fibre and expensive on capped LTE. If you are rolling this out to a team, mirror the packages locally and stage the ISO once.

Where it fits our delivery model. This is a developer-workstation decision, not a client-platform decision. It belongs in the same conversation as owning your source and your configs — a machine whose entire state is text in Git is the logical endpoint of that argument. It does not belong in a proposal for a client endpoint estate, and pitching it there will cost you credibility on the parts that matter.

Where this links in the tree.

The workstation is the bottom of the stack the rest of the tech branch sits on. These are the nodes it touches directly.

Primary sources.

Project sites, the source repository, the release notes and DHH's own essays. Content validated August 2026 against Omarchy 4.0.0 (released 14 August 2026).