Deskwright

Computer use for AI agents on GNOME Wayland.

Let your coding agent use your Linux desktop. Or give it one of its own, so it stops stealing your mouse.

Install  •  What it can do  •  A desktop of its own  •  When something's off  •  Security

Claude Code can edit your files and run your commands. It can't open GIMP, click a button in a settings dialog, or read what a native app is showing you. On macOS and Windows it can, through computer use. On Linux that's still on Anthropic's list.

This fills the gap on GNOME. It's an MCP server plus a small shell extension, and it hands an agent the actual desktop: launching apps, reading widgets, clicking, typing, dragging, window management, OCR and screen recording. Any MCP client can drive it. Claude Code, Codex, Cursor, your own script.

The part people tend to like most: DESKWRIGHT_SESSION=headless runs all of it on a virtual monitor that isn't on any of your screens. Your agent gets a real GNOME desktop to work on, and it never takes your focus.

Why this exists

Wayland deliberately stops an application from seeing or touching any other window. That's a good rule, and it's why xdotool, wmctrl and grim either do nothing on GNOME or fail in ways that look like your own mistake. Every Linux computer-use project I found either shipped a Docker container running X11, or quietly assumed X11 and broke.

So I went looking for what GNOME actually permits, and it turns out to be quite a lot, just not where anyone looks. Mutter answers D-Bus. AT-SPI, the accessibility layer built for screen readers, exposes every real widget in every running app with its name, its role and the action it performs. And a shell extension runs inside gnome-shell itself, where the rest lives.

That middle one matters more than it sounds. Pressing a widget's own accessibility action isn't a nicer way to click. It's a different thing: it can't miss, it survives the window moving, and it needs no pointer at all. An agent working this way stops guessing at coordinates, and stops taking a screenshot after every action to find out what happened.

Install

You need GNOME Shell on Wayland. Check with echo $XDG_CURRENT_DESKTOP $XDG_SESSION_TYPE, which should mention GNOME and wayland.

The lazy way: hand it to your agent

Open Claude Code, or Codex, or whatever you use, and say:

Set up https://github.com/tristanmuzzu/deskwright on this machine, follow the AGENTS.md.

AGENTS.md is a runbook written for agents. Every command, how to check each one worked, the right package names for Debian, Fedora and Arch, and the handful of things that go wrong with their fixes. Your agent will ask you for a sudo password once and tell you to log out once. That's your whole involvement.

By hand: two commands and a logout

pipx install --system-site-packages deskwright
deskwright-setup

deskwright-setup narrates every step. It turns on the accessibility flag, installs the bundled shell extension, and tells you about any missing system package with the right install line for your distro, so you're never guessing at package names. It never runs sudo itself.

Then log out and log back in. Once. There's no way around this one: on Wayland, gnome-shell only picks up an extension at session start.

Last step, point your client at it:

claude mcp add deskwright --scope user -- deskwright

That's it. Ask your agent to open an app and it will.

claude plugin marketplace add tristanmuzzu/deskwright
claude plugin install deskwright@deskwright

~/.claude/plugins/marketplaces/deskwright/bin/deskwright-setup

git clone https://github.com/tristanmuzzu/deskwright
cd deskwright
bin/deskwright-setup
./mcp_server.py --self-test

Check it works

DESKWRIGHT_SESSION=headless deskwright --self-test

You want 18/18 passed. It runs on a virtual monitor rather than your screen, so it's safe to run while you're working. The first run takes about 20 seconds because it has to start a second gnome-shell.

Drop DESKWRIGHT_SESSION=headless and it tests your real desktop instead. Do that one while you're looking at the screen: the self-test injects real input, because it's checking the guards that refuse dangerous key combinations.

When something's wrong, ask your agent to call desktop_health. It answers in one line whether this desktop is usable, then says which mechanisms work right now and what each of them will actually do.

What it can do

33 tools, roughly 11k tokens of schema in a session. That's the honest price, and it's why each one returns enough that you don't need a second call to work out what happened. The order below is the order an agent should reach for them. Accessibility tree first, pixels last.

A second desktop it uses while you work

An agent that needs your screen is only half useful. deskwright-headless starts a separate GNOME session on a virtual monitor, with its own session bus, its own gnome-shell --headless and its own runtime directory. A server pinned to it drives that desktop with the same 33 tools while you keep the physical one.

deskwright-headless start                    # about 200 MB of gnome-shell, idempotent
deskwright-headless status                   # liveness, memory, bus address
deskwright-headless stop                     # don't leave it idling on an 8 GB machine

Register it as a second MCP server and you can hand it long jobs:

claude mcp add deskwright-headless --scope user --env DESKWRIGHT_SESSION=headless -- deskwright

Sessions are named, so two agents can each have a desktop of their own and never watch each other's windows move:

deskwright-headless start --name work
deskwright-headless list                     # every session, memory used, memory free
DESKWRIGHT_SESSION=headless:work deskwright

There are guards on this, because each session is a real compositor at around 205 MB: a per-name start lock so two agents can't both spawn one, a session cap (DESKWRIGHT_HEADLESS_MAX, default 4), and a free-memory floor that refuses a start which would push the machine into swap. The agent that would cause that can't see it coming, so the server does.

When something's off

The extension says INACTIVE, or window tools don't work. You haven't logged out yet. gnome-shell can't load an extension without a session restart and there's no workaround. Until then you still get AT-SPI, pointer, keyboard and clipboard. You don't get window management, extension screenshots, pointer position or the halt switch.

An app shows no widgets in ui_tree. It was already running when toolkit-accessibility got turned on. Apps read that setting at startup, so restart the app.

Chrome, Electron or Qt apps look empty. As far as AT-SPI is concerned, they are. Use find_text instead, which OCRs the screen and hands back coordinates. For Chrome specifically, launching it with --force-renderer-accessibility gets you a real tree at a small performance cost.

Everything fails with input_backend_failed. pipx without --system-site-packages. Reinstall with the flag:

pipx install --force --system-site-packages deskwright

Every ui_* call says "Namespace Atspi not available". You have python3-gi but not the AT-SPI typelib, which is a separate package: gir1.2-atspi-2.0 on Debian and Ubuntu, at-spi2-core on Fedora and Arch. deskwright-setup --check catches this and names it.

A tool returns halted. Somebody pressed Super+Ctrl+Escape, which is the halt switch. Press it again to clear it.

Your screen locked and half the tools stopped. Expected. GNOME unloads extensions that don't declare unlock-dialog, so screenshots and window geometry go away until you unlock. AT-SPI keeps working, so ui_find, ui_press and ui_read_text all still do.

What it won't do

Deliberately out of scope as a project, not just off by default:

• No CAPTCHA solving. CAPTCHAs exist to tell humans from software. This is software.
• No credential typing features. Nothing here is built to harvest, store or autofill secrets. The journal doesn't record what gets typed, only how much.
• No detection evasion. No timing jitter to look human, no fingerprint spoofing, no anti-anti-bot work.
• No cloud, no telemetry, no account. It's a local process on your session bus, and nothing leaves the machine.

On the other side of that line the design is deliberately permissive. The agent is trusted, and the tooling's job is to make it capable rather than to fence it in. The whole safety budget goes to three things: a halt switch a human can always reach, guards against actions that can't be undone, and an honest record of what happened. SECURITY.md is blunt about what enabling this actually switches on, and it's worth reading before you point it at your real screen.

How it actually works

Wayland denies all of this to Wayland clients. It says nothing about D-Bus, and that's where the doors are. Four mechanisms, roughly in order of how much work they carry:

1. AT-SPI for anything semantic. Real widgets, real actions, no pointer.
2. A gnome-shell extension over D-Bus for what gnome-shell keeps to itself: window enumeration and control, screenshots, pointer position, and a keybinding grab for the halt switch.
3. org.gnome.Mutter.RemoteDesktop for pointer and keyboard. Absolute coordinates, and keysyms rather than key positions, so your keyboard layout can't transpose what gets typed.
4. xdg-desktop-portal for the same thing, standardised. This is the route to compositors that aren't GNOME.

Worth knowing: number 2 is the only one of those that isn't already open. Mutter's ScreenCast and RemoteDesktop interfaces answer any client on your session bus with no consent dialog, which is why recording here needs no permission popup, and why the portal is a caller of them rather than a gate in front of them. SECURITY.md has the exact commands if you'd rather check that yourself than take my word for it.

docs/field-notes.md is the long version: every wall hit on the way here, what the compositor refused, and the measurements that shaped the tool surface. It's the file to read when something behaves oddly, or before you change the code.

Requirements

deskwright-setup --check is the real answer. It detects everything, names the package for your distro and exits nonzero if a hard requirement is missing. The short version:

Claude Code, not Cowork. On Linux the desktop app runs Cowork tasks inside a QEMU/KVM virtual machine, so an MCP server it starts lives in that VM and cannot see your session bus, your gnome-shell or your screen. Everything here would fail there. Use it from the Claude Code CLI, the Code tab, or any other MCP client running on the host itself.

Claude Code, not Cowork. On Linux the desktop app runs Cowork tasks inside a QEMU/KVM virtual machine, so an MCP server it starts lives in that VM and cannot see your session bus, your gnome-shell or your screen. Everything here would fail there. Use it from the Claude Code CLI, the Code tab, or any other MCP client running on the host itself.

Packages, using Debian names (deskwright-setup prints yours): python3-gi, gir1.2-atspi-2.0, python3-pil, libglib2.0-bin, wl-clipboard, tesseract-ocr. ydotool is optional and only used as an input fallback. The headless session additionally wants gnome-shell and dbus-daemon as binaries, which any normal desktop already has.

Contributing, security, license

CONTRIBUTING.md has the layout, the test suites and what each one needs. SECURITY.md is the threat model, stated plainly. ROADMAP.md is what's next and why, in order.

Apache-2.0. If you get it running on a compositor that isn't GNOME, or on a GNOME older than 50, please open an issue and say so. That's the most useful thing anyone can send.

_{mcp-name: io.github.tristanmuzzu/deskwright}