Claude Desktop Comes to Linux: Cowork Runs a Real VM Inside Your PC

On June 30, 2026, Anthropic, the company behind the AI assistant Claude, released a Linux build of its desktop app. What used to be macOS- and Windows-only now runs officially (as a beta) on Ubuntu and Debian.

To be honest, the "how to install it" steps no longer hold much value. You can just ask Claude itself and it will walk you through them. What this article is actually about is something else: what makes the desktop Claude fundamentally different from Claude in the browser, and the one thing on Linux that almost everyone trips over — the "virtualization permission" pitfall. I ran into the latter on an actual Linux machine, so I'll cover that too.

Here is the conclusion up front: the headline feature, "Cowork," actually spins up a full second Linux computer (a virtual machine) inside your own PC, and runs its shell commands and code in there. That is why it can safely touch your files. That is why Linux needs a "permission to use virtualization." It all connects.

First: what the Linux build can and can't do

Let me get the scope out of the way in one paragraph. This Linux build is a beta. Supported systems are Ubuntu 22.04 or later / Debian 12 or later (both Intel-family x86_64 and Arm64). You install it from an apt repository (the standard software-management mechanism on Ubuntu/Debian), so future updates arrive through normal system updates. Fedora and RHEL are not supported yet, with more distributions said to be planned.

The interface has all three of "Chat," "Cowork," and "Code." On the other hand, two features that exist on macOS/Windows are disabled on Linux.

"Computer Use" is the feature where Claude looks at your screen and operates the mouse and keyboard. It's disabled on the Linux build for now. That covers the facts. From here on is the real subject.

What makes it fundamentally different from Claude in the browser

The first question is: "If I can just open claude.ai in a browser, why bother installing an app?" The answer is clear-cut: the desktop app can touch your PC itself.

Browser Claude basically just talks. It can't read your folders, rewrite your files, or run commands. The desktop app, by contrast, can actually operate your local machine (the very PC you're using right now) through three routes.

• 1 MCP (a common mechanism for connecting Claude to external tools and files) lets it read and write files in the folders you designate
• 2 Cowork autonomously handles multi-step work such as organizing files, creating documents, and processing data
• 3 Computer Use operates apps by looking at the screen (disabled on the Linux build)

If you assume the browser and desktop versions are "the same Claude," you'll miss this gap. Cowork is desktop-app only to begin with — it doesn't exist in the browser. The real substance of this news is that this desktop-only feature now runs on Linux too. Cowork is available on the Pro (US$20/month), Max, Team, and Enterprise plans.

What Cowork really is: "spinning up another Linux inside your PC"

This is the single most important point of the article. When Cowork touches your files or runs a command, it is not executing that directly on your PC. According to Anthropic's official architecture overview, Cowork boots a full, isolated, real Linux virtual machine (VM) inside your PC and runs shell commands and code in there.

A virtual machine (VM) is "another independent computer" created in software inside a single PC. Whatever happens inside the VM is cut off from your actual PC (the host) on the outside. Since you're letting an AI run arbitrary commands, this isolation makes sense. Even if Claude goes off the rails or accidentally runs a dangerous command, the damage is confined to the VM. On top of that, only the files in the folders you've allowed are passed between the VM and the host.

What's interesting is that the "software that runs the VM (called a hypervisor)" is a completely different one on each OS.

macOS uses Apple's own virtualization stack, Windows uses Microsoft's Hyper-V, and both work without the user doing anything special. But Linux alone requires that Claude's app be granted direct permission to use "KVM," the virtualization feature built into the OS. This is where people get stuck.

What actually appeared: "Virtualization isn't fully set up"

When I installed it on a Linux machine and tried to launch Cowork, this message appeared.

Virtualization is not fully set up.
Cowork runs on a secure local virtual machine.
Claude does not have permission to use virtualization (/dev/kvm).
Add your user to the 'kvm' group,
then log out and log back in.

The previous section shows up verbatim on screen. /dev/kvm is the entry point (a device file) to the "hardware for running virtual machines fast" that Linux ships with by default. Cowork's VM runs through this, but out of the box a regular user isn't granted permission to use it. So you're prompted to "add yourself to the kvm group so you can access that device."

The fix is exactly as the message says: run this one line, then log out and back in.

# Add yourself to the kvm group ($USER expands to your username)
sudo usermod -aG kvm $USER

# You must log out and back in for it to take effect (a reboot also works)

Group membership doesn't apply to the session you currently have open. People tend to skip this and then puzzle over "it still doesn't work," so don't forget to log out and back in (or reboot). Conversely, that's pretty much the only place you get stuck. The steps are simple, but if you hit that abrupt, partly-English error without knowing "why the kvm group is needed," your hands freeze.

It doesn't "run anywhere." No KVM, no launch

There's a deeper constraint than permissions. The official Linux build assumes that the KVM virtualization feature is available in the first place, and it provides no software substitute (emulation) for environments that lack it. In other words, it's not a permission problem — Cowork simply cannot launch in an environment where KVM doesn't exist.

Concretely, the representative cases are:

• • Cloud Linux servers where "running a VM inside a VM" (nested virtualization) is disabled
• • Container environments (Docker, etc.) where the KVM device hasn't been passed through

Cowork boots this VM internally even just to run a single shell command. In an environment where it can't stand up a VM, it stops right at the door. If you're thinking of "installing the app on a remote cloud Linux box and running Cowork there," check whether KVM is usable on that machine first to avoid a wasted trip. For use on your own PC (a physical machine, or one with virtualization support enabled), the kvm group setup from the previous section is all you need.

For reference, Cowork on Linux is built by combining well-established virtualization parts — QEMU (software that runs virtual machines), OVMF (boot firmware for the VM), and virtiofs for sharing folders between the host and the VM. It's not some flashy new technology; it's existing pieces cleanly folded into a desktop app.

So who can use it, and who gets stuck

To sum up. If you're on Ubuntu 22.04+ / Debian 12+, on your own PC with KVM enabled, and on a paid Pro-or-higher plan, then Linux Claude Desktop — including Cowork — works after just one extra step of usermod -aG kvm and logging back in. Where people get stuck is mostly two patterns: trying to run it on a cloud or container without KVM, and skipping the log-back-in after the kvm group change.

As I said at the top, the install steps themselves are something you can just ask an AI about these days. What has value is knowing "why it's designed this way" and "where the pitfalls are" ahead of time. Once you grasp the single fact that Cowork stands up a real Linux VM inside your PC and touches real files in that isolated space, the "virtualization permission" error on Linux and the reason it won't launch in the cloud all connect as the same story.

On the Claude side, it's been a turbulent stretch — the release of the top model "Claude Fable 5," the story of its suspension and return, and the leak of Claude Code's source. The Linux support for the desktop app is quiet by comparison, but it's a practically useful step in turning AI from a "conversation partner" into a "hands-on collaborator on your own machine."

Caveats for this article

Because the Linux build is a beta, the supported systems, behavior, and error wording may change. The error and fix described here (adding to the kvm group and logging back in) reflect what I confirmed by actually installing it on an Ubuntu-family machine. The internal virtualization details (QEMU, virtiofs, etc.) are based on the official architecture overview and outside technical analysis; specifics may differ by environment and version.

Sources

• • Download Claude (Anthropic official app download page)
• • Claude Cowork desktop architecture overview (Anthropic Help Center — per-OS hypervisors)
• • Get started with Claude Cowork (Anthropic Help Center)
• • Anthropic releases Claude Desktop app beta for Linux users (Crypto Briefing — scope and beta limits)
• • Inside Claude Cowork: How Anthropic Runs Claude Code in a Local VM (technical analysis — the local VM internals)
• • Securing Claude Cowork: A Security Practitioner's Guide (Harmonic Security — the VM-isolation angle)