Umbriel

umbriel-mcp

mcp-name: io.github.dot-ASH/umbriel-mcp

An MCP server for the Umbriel Wayland compositor.

Exposes Umbriel to MCP clients: list and control windows, workspaces and outputs, drive the pointer and keyboard, capture the screen with optional OCR, and read or write the clipboard.

Compositor state and actions go straight to Umbriel's IPC socket rather than shelling out to the CLI, so inspection costs one round trip and no process spawn. Pointer, keyboard, and clipboard use Wayland protocols directly. Only screen capture and OCR call external programs.

Read this before installing

Full access means full control of your desktop. In the default full profile an agent connected to this server can launch arbitrary commands, type and click anywhere, read and overwrite your clipboard, and see your screen. Anything on screen or in the clipboard is visible to the model. Use the profiles to grant less.

This is a best-effort personal project that tracks an actively evolving compositor. Expect 0.x behavior changes, and expect it to break when Umbriel changes. There is no support promise beyond what I can do in my spare time.

Compatibility

Support follows Umbriel's own availability: if Umbriel runs on your system, this server should. The server itself is a static binary for x86_64 and aarch64 Linux with no runtime libraries, so the distribution only matters for installing Umbriel and the two screen helpers.

Older Ubuntu releases are not supported because Umbriel is not packaged for them.

Warning for Arch users: the AUR package named umbriel is a placeholder and is not the compositor. Do not install it; the compositor is umbriel-git.

Install

Each path below installs Umbriel, the screen helpers, and umbriel-mcp, then checks the result. Skip the Umbriel step if you already run it. The Umbriel commands are copied from its installation page; check there if they have changed.

Arch Linux

yay -S umbriel-git                                     # the compositor, if neededsudo pacman -S grim tesseract tesseract-data-eng wf-recorder   # screenshots, OCR, recording

Then install the server as a pacman package built from this repository's recipe. It downloads the prebuilt release binary and checks its checksum, so nothing is compiled:

git clone https://github.com/dot-ASH/umbriel-mcpcd umbriel-mcp/packaging/aur-binmakepkg -siumbriel-mcp doctor

Remove it later with sudo pacman -R umbriel-mcp-bin. An AUR package of the same name is planned; this section will switch to yay -S umbriel-mcp-bin once it is published.

Ubuntu 26.04 and newer

1. Install Umbriel from the Noctalia APT repository (amd64 and arm64 only), if needed:

   sh

   wget https://pkg.noctalia.dev/deb/nickh-archive-keyring.debsudo dpkg -i nickh-archive-keyring.debsudo wget -O /etc/apt/sources.list.d/noctalia-resolute.sources \  https://pkg.noctalia.dev/deb/noctalia-resolute.sourcessudo apt updatesudo apt install umbriel

2. Install the screen helpers:

   sh

   sudo apt install grim tesseract-ocr tesseract-ocr-eng wf-recorder

3. Install the server:

   sh

   curl -fsSL https://raw.githubusercontent.com/dot-ASH/umbriel-mcp/main/install.sh | sh

   This downloads the latest release for your architecture, checks it against the release's SHA256SUMS, installs it to ~/.local/bin/umbriel-mcp (no sudo), and runs umbriel-mcp doctor. Ubuntu adds ~/.local/bin to PATH at login once the folder exists, so log out and back in if umbriel-mcp is not found afterwards. If you installed an older version into /usr/local/bin, the script tells you to remove it.

   Manual install without the script

   sh

   arch=$(uname -m)   # x86_64 or aarch64tag=v0.4.1base="https://github.com/dot-ASH/umbriel-mcp/releases/download/$tag"curl -LO "$base/umbriel-mcp-$tag-$arch-linux.tar.gz"curl -LO "$base/SHA256SUMS"sha256sum --check --ignore-missing SHA256SUMStar -xzf "umbriel-mcp-$tag-$arch-linux.tar.gz"install -Dm755 "umbriel-mcp-$tag-$arch-linux/umbriel-mcp" ~/.local/bin/umbriel-mcp

4. Check the setup from a terminal inside the Umbriel session:

   sh

   umbriel-mcp doctor

Debian Trixie and Sid

Same as Ubuntu, with the Debian repository source and, on Trixie, the backports suite:

wget https://pkg.noctalia.dev/deb/nickh-archive-keyring.debsudo dpkg -i nickh-archive-keyring.debsudo wget -O /etc/apt/sources.list.d/noctalia-trixie.sources \  https://pkg.noctalia.dev/deb/noctalia-trixie.sources        # Sid: noctalia-unstable.sourcessudo apt updatesudo apt install -t trixie-backports umbriel                  # Sid: sudo apt install umbrielsudo apt install grim tesseract-ocr tesseract-ocr-eng wf-recorder

Then install the server, which also runs umbriel-mcp doctor:

curl -fsSL https://raw.githubusercontent.com/dot-ASH/umbriel-mcp/main/install.sh | sh

Fedora

sudo dnf install umbriel-nightly                              # from the Terra repositorysudo dnf install grim tesseract tesseract-langpack-eng wf-recorder

Then install the server, which also runs umbriel-mcp doctor:

curl -fsSL https://raw.githubusercontent.com/dot-ASH/umbriel-mcp/main/install.sh | sh

Other distributions

Install Umbriel following its installation page, then grim, tesseract, and Tesseract's English data (eng.traineddata) from your package manager. Install the server with curl -fsSL https://raw.githubusercontent.com/dot-ASH/umbriel-mcp/main/install.sh | sh, or build it with Cargo (needs Rust 1.89 or newer, for example from rustup):

cargo install umbriel-mcp --locked

umbriel-mcp doctor names the exact packages to install on Arch, Debian/Ubuntu, and Fedora, and gives generic guidance elsewhere.

Updating

umbriel-mcp update            # install the latest release in placeumbriel-mcp update --check    # only report whether one exists

update downloads the latest release, checks it against SHA256SUMS, confirms the new binary runs and reports the expected version, and only then replaces the old one, so a failed update leaves the old binary working. Restart your harness afterwards; running sessions keep the old version. umbriel-mcp doctor also says when a newer version exists. Nothing checks for updates in the background.

• If the binary is in a folder you cannot write to, such as /usr/local/bin, it asks you to run sudo umbriel-mcp update.
• If it was installed by a package manager (pacman or makepkg, cargo install, Nix), it refuses and tells you the package manager's command instead, so the two never disagree.
• Versions before 0.3.0 cannot update themselves to a signed release; run the install script once to get 0.3.0, and umbriel-mcp update works from then on.
• Releases are signed. update and the install script accept a release only if its SHA256SUMS carries a valid signature from the umbriel-mcp release key, whose public half is built into the binary and the script, and only then trust its checksums. A tampered or unsigned release is refused. Updates are protected even if the GitHub account were compromised; a first install with the script is not, because the script itself comes from GitHub.
• Releases 0.2.3 and older are unsigned, so the script and update only install 0.3.0 and newer.

Uninstalling

umbriel-mcp uninstall

This lists what it will remove and asks once (--yes skips the question):

• the umbriel server registered with pi, Claude Code, opencode, Cursor, VS Code or Zed. Claude Code is unregistered through its own CLI; in the others' config files only the umbriel entry is deleted, and comments, formatting and other servers stay as they were. An entry named umbriel that runs some other program is left alone.
• the umbriel-mcp program itself, when the install script or umbriel-mcp update put it there. If you run it through a symlink, only the link is removed. For a pacman, cargo or Nix install it prints the command to run instead.

--keep-binary only unregisters. Config backups and screen recordings are listed but kept. Restart any running harness afterwards. Registrations inside a single project (a project's opencode.json or .cursor/mcp.json) are not searched; delete those entries by hand.

Requirements

Pointer, keyboard, and clipboard need nothing beyond a running Umbriel session. These external programs are needed only for the screen tools:

The tools are always listed; a call that needs a missing program fails with an error naming the package to install.

The socket is found through UMBRIEL_SOCKET, falling back to $XDG_RUNTIME_DIR/umbriel-$WAYLAND_DISPLAY.sock.

Register with your harness

umbriel-mcp install pi            # also: claude-code, vscode, opencode, cursorumbriel-mcp install zed           # prints config to pasteumbriel-mcp install claude-code --profile no-exec

This registers the server through each harness's own CLI, so umbriel-mcp never reads or edits their config files itself:

If opencode's or Cursor's CLI is not installed, or Cursor's has no mcp add, the config to paste is printed instead. A profile is added to the registration only if you pass --profile, so the default stays full.

The pi registration uses direct exposure, so the model sees the tools in its tool list and answers "what's on workspace 4?" with one list_windows call instead of searching for tools or trying hyprctl. If you registered pi by hand earlier, run umbriel-mcp install pi again to switch to direct exposure. The Zed snippet keeps "args": [] because Zed rejects its whole settings file if it is missing.

Any other MCP client can use the plain stdio form:

{  "mcpServers": {    "umbriel": { "command": "umbriel-mcp", "args": [] }  }}

Checking your setup

umbriel-mcp doctor verifies that the Umbriel socket is reachable, that the compositor offers the virtual-pointer and virtual-keyboard protocols, and that grim, tesseract, and Tesseract's English data and wf-recorder are installed. Each problem is listed on its own with the command to fix it on Arch, Debian/Ubuntu, and Fedora, or generic advice on other distributions. It exits with status 0 only when everything is ready.

Profiles

--profile limits what the server will do. Every tool stays visible in every profile; a call the profile does not allow returns an error that names the profile.

no-exec stops the server from starting programs; it is not a sandbox. An agent that can type into a focused terminal can still run commands through the input tools.

Tools

Inspection — list_windows, list_workspaces, list_outputs, get_active_window, list_layers. Window and monitor listings are brief by default: id, app, title, place, size, workspace and has_keyboard_focus for windows; name, position, size, current mode, scale and workspace for monitors. Pass detail: true for every field Umbriel reports, including every display mode.

Window and workspace control — switch_workspace, focus_window, close_window, kill_window, move_window_to_workspace, toggle_fullscreen, toggle_floating, resize_window, set_workspace_layout, toggle_overview, launch_app, run_action

Input — mouse_move, mouse_click, mouse_scroll, mouse_drag, type_text, key_press. These drive your real mouse and keyboard, so they collide with you if you use the desktop at the same time; the server tells agents to ask before taking over. To keep input where it was meant to go:

• type_text and key_press take a window_id: if that window does not have keyboard focus, nothing is typed and the error names the window that does.
• mouse_click and mouse_move take a window_id with x and y counted from that window's top-left corner. A window that is not on screen, or a point outside it, is refused.
• After typing, pressing or clicking, the result says which window has keyboard focus.

About Umbriel itself — umbriel_info (version, keyboard layouts, submap, color and tearing state), list_actions (every action run_action and keybinds accept, filterable), config_options (exact config keys with types, defaults and allowed values, from Umbriel's own schema), read_config (your config or one section such as keybinds, plus validation). These read from the installed umbriel command and your config, so they match the version you run. The schema has no prose descriptions; explanations live in the Umbriel documentation.

Changing settings — set_option and reset_option change one setting, such as layout.gap or appearance.blur.enabled, and set_keybind and remove_keybind change keybinds. Keys and values are checked against Umbriel's schema (type, range, allowed values), then the edited config is checked with umbriel config validate. Rule lists such as window_rule are edited with add_rule, replace_rule and remove_rule, which take a rule as JSON and a position counted from 1 in the order read_config shows.

Every change needs your approval before anything is written:

• In clients that support MCP elicitation (Claude Code 2.1.76+, VS Code, Cursor, opencode v2), you get a dialog showing the exact change, with a checkbox for keeping a backup. Only your approval writes it; the model cannot answer the dialog. No answer within 10 minutes writes nothing.
• In clients without it (pi, opencode 1.x), the tool writes nothing and returns the change with a one-time confirmation code. The model is told to show you the change, ask whether to keep a backup, and wait for your answer before calling apply_config_change. This relies on the model following that instruction, so it is weaker than a dialog. All write tools are marked destructive, so a harness permission prompt or extension can ask you as well.
• A code expires after 10 minutes, works once, and is refused if the file changed since the preview.

The same approval step asks whether to keep a backup of the previous config. If you say yes, or don't answer that question, it is saved next to the config as config.toml.bak-<time>; say no to avoid collecting backup files. umbriel-mcp keeps the newest 5 of its own backups and deletes older ones whenever it writes a new one; start it with --keep-backups N to keep more (up to 1000). When a change is written, only the changed lines differ (comments and formatting are kept), a symlinked config stays a symlink, and Umbriel applies it immediately.

Profiles still apply. read-only refuses all changes. no-exec refuses spawn: keybinds and hot-corner actions, and settings that can start programs, load files, or weaken a security check (general.autostart, events, environment, include, effect shaders, drm, screencast.disable_dynamic_confirmation); those need full. Disabling your only enabled output is always refused, and keyboard-layout or general.mod_key changes come with a warning.

Recording — start_recording records one monitor, or the screen area of one window that is on screen, to ~/Videos/umbriel-mcp/recording-<UTC time>.mp4 (or your XDG_VIDEOS_DIR). Audio is off unless asked for. A recording stops by itself after 30 seconds by default (1 second to 60 minutes allowed), stop_recording stops it sooner and reports the file, and only one runs at a time. If umbriel-mcp exits, the recording is stopped and the file finished. Recording is observation like screenshots, so every profile allows it. wf-recorder shows nothing on screen while it records.

Choosing what to capture — when you ask for a screenshot or a recording without saying what to capture, you are asked to choose: the whole screen (screenshots only), a monitor, or a window. For a recording you are also asked how long, from 30 seconds to 60 minutes, with 30 seconds preselected, unless you already said how long. Clients with dialogs show the choices in one; others have the model ask you. When the model takes a screenshot or starts a recording for its own purposes, it does not ask and never waits for you.

Screen and clipboard — screenshot, screenshot_with_ocr, find_text_on_screen, click_text, wait_for_text, wait_for_change, clipboard_read, clipboard_write.

• With a window_id, captures and text search cover only that window, even on a hidden workspace, without switching to it (needs grim 1.5 or newer; older grim can only capture windows that are on screen). This leaves your other windows out of the picture.
• With a region (x, y, width, height in screen coordinates), they cover only that rectangle, at full resolution, to zoom in on a detail.
• Every capture says when it was taken, and notes when a window's title changed since its last capture, which usually means it shows a different page or someone else is using it.
• wait_for_text waits until text appears. wait_for_change waits until a window, monitor or region differs from your previous capture of it and then stops changing, and returns the settled picture; if nothing changed, it says so. Both give up after 10 seconds by default (30 at most), so an agent can wait for a page to load instead of guessing with sleeps.

key_press takes chords such as ctrl+shift+Tab, super+q, F5 or alt+Left; an uppercase letter (ctrl+A) implies Shift. type_text accepts any Unicode text, with \n typed as Return and \t as Tab.

run_action accepts any action from umbriel msg --help, so the compositor's full verb set stays reachable even where there is no dedicated tool.

Notes

Workspaces in Umbriel belong to an output rather than to a global pool. Selectors take a position or a name and accept an optional /output suffix, which the workspace and output arguments build for you.

Screenshots are downscaled to 1024px wide by default and returned as JPEG alongside a coordinate-mapping note, so screen coordinates can be recovered from image coordinates. OCR is different: it always runs on the full-resolution capture, because recognising text in the downscaled preview finds nothing. find_text_on_screen and click_text therefore report absolute screen coordinates directly, and match multi-word phrases such as Open File.

License

MIT