deskpilot

Desktop control for agents on Hyprland. An agent takes a screenshot of a window, picks a point in it, and clicks or types there, in any app: native Wayland, XWayland and Wine programs such as the Battle.net launcher and games.

Deskpilot never blocks the user. Clicks, scrolls, drags and keys run inside the compositor through the deskpilot Hyprland plugin, which sends them straight to the target app's own pointer and keyboard: the user's pointer, cursor, keyboard focus and input are left alone while they keep working, and the target is not raised, even with focus_on_activate set. Two cases cannot work that way, and there the plugin briefly holds the user's mouse input, moves the seat's pointer onto the window, presses, and puts the pointer back, applying the motion the user made meanwhile: an XWayland window (X11 and Wine apps, which read the one X pointer), and a window of the app the user's pointer is on. Keys likewise hold the user's input for a stroke only when their keyboard focus is in the target's app or the target is an XWayland window. In those cases the plugin refuses a click when another window covers the point.

An agent cursor shows where deskpilot acts. The plugin draws it above every window, separate from the user's pointer: it glides along an arc to each point before the action starts, shrinks and sends out a ring at each press, follows a drag as it moves, and fades after a few idle seconds. Window screenshots never include it. DESKPILOT_CURSOR=off turns it off.

--hidden acts on a window wherever it is, on a hidden or special workspace included, fully out of the user's view: no workspace switch, no agent cursor, and the user's pointer, focus and screen stay as they are. The plugin sends pointer input straight to that window's own client, and lets it draw while off screen so a screenshot after the action shows the result. Tooltips and menus it opens show in that screenshot and take its clicks, and they never grab the user's input. It works for native Wayland windows. An XWayland window (X11 and Wine apps) is refused, since X clients read the one X pointer only the user's pointer moves; so is a window of the app the user's pointer is on, and a drag and drop the press starts is cancelled.

Keys sent to a window (--window) go to that window only, with each stroke complete before focus returns, and the user's modifiers never leak into them. X11 apps take keys only from the window with X input focus, so an XWayland target gets X focus for each stroke and sees a focus-in and focus-out around it; Hyprland's own focus does not move.

Text is typed as written on the user's own keymap, whatever its layout and with Caps Lock on: the plugin finds the key that gives each character on the active keyboard's keymap, which every client has, and sets the modifiers it needs directly. A character no key on that keymap gives is an error rather than skipped.

--take-pointer uses a virtual mouse made with /dev/uinput instead, which the compositor treats as real hardware. Tools that synthesize input another way, such as xdotool (XTest), are ignored by some clients: Battle.net drops xdotool clicks.

Requirements

• Hyprland, with hyprctl on PATH. Version 0.57's Lua dispatchers and the older string ones both work.
• The plugin build tools: a C++26 compiler, make, Hyprland's headers (pkg-config hyprland, shipped with Arch's hyprland package), libxcb, libxkbcommon and cairo.
• grim for screenshots.
• wl-clipboard for clipboard.
• For --take-pointer: write access to /dev/uinput, usually through the input group or a udev ACL. Check with test -w /dev/uinput.
• Python 3.12 or later and uv.

Install

uv tool install deskpilot

The agent skill, which teaches an agent the steps below, installs in Claude Code with:

/plugin install deskpilot --marketplace cjber/deskpilot

Use

deskpilot windows                         # every window: address, class, title, geometry, whether on screen
deskpilot shot battle.net.exe             # screenshot; prints its id, image path and scale
deskpilot click battle.net.exe 920 367 --shot ID    # a point in that screenshot
deskpilot drag gimp 100 400 500 400       # press, move and release: sliders, selections, drag and drop
deskpilot hover firefox 300 40 --wait 1   # rest the pointer there; prints a shot showing the tooltip
deskpilot elements gedit                  # the app's accessibility tree, with ids
deskpilot click gedit --element e12       # the centre of that element, where the app says it is now
deskpilot key Right --window kitty --hold 2         # hold a key down two seconds
deskpilot type "/reload" --window wow.exe --delay 0.12
deskpilot key ctrl+a --window kitty
deskpilot paste "Grüße, 日本" --window firefox    # any text, through a clipboard only that window sees
deskpilot clipboard                       # the user's clipboard text; --set TEXT replaces it
deskpilot wait-for wow.exe --timeout 120
deskpilot wait-for wow.exe --changed ID --region 0 0 400 80   # until that part of shot ID looks different

A window is named by its address, its class, or a substring of its class or title. A name that matches more than one window is an error listing them, so a click never lands on a guessed window.

shot captures the window itself, so a window on top does not show in it, and saves it under an id. move, click and scroll take pixels in the shot named by --shot, or the window's latest one, whatever scale it was taken at. A window resized since that shot is refused, as is a point outside the window. Pass --scale 1 to give window pixels instead.

A window on a hidden workspace or scrolled off its monitor is reported as "visible": false, and shot, move, click, scroll, drag, hover and elements refuse it rather than move the user's view. --hidden acts on it where it is, out of view; --show focuses it to bring it on screen first.

The first action that needs the plugin builds and loads it for the running Hyprland, so it comes back by itself after Hyprland restarts or upgrades. deskpilot plugin load does the same ahead of time.

Some apps close a menu or dropdown as soon as it opens when their window is not the focused one, as Chrome does with many web page dropdowns. click --take-focus focuses the window first and leaves it focused, so the list stays open for the next click.

move only shows the agent cursor. Pass --take-pointer to move, click or scroll to move the user's own pointer there and press through the input device instead.

drag presses at one point, moves to another over --duration seconds (0.3 by default, up to 5) and releases, with the user's pointer left where it is: sliders, text selection, drawing, and drag and drop within the window. hover rests the pointer on a point for --wait seconds and prints a screenshot taken while it is there, so a tooltip or a hover state shows in it; --click X Y clicks a second point next instead, before the pointer leaves, for a menu that opens on hover. Both are bounded, and a drag or hover that fails, or whose window closes, still lets go of its button.

elements prints the window's accessibility tree (AT-SPI): each element's id, role, name, states, text and bounds in window pixels. click, move, scroll and hover take --element ID instead of a point, and act on the element's centre where the app reports it at that moment. Each elements call prints a snapshot id, and SNAPSHOT:e5 pins an id to that snapshot: it is refused as stale once a newer snapshot replaces it, since every snapshot numbers its elements afresh. An element whose role or name changed since is stale too, and one no longer showing is offscreen. elements --peek reads the tree without replacing the latest snapshot, so a program can check what changed before it acts on an id it already holds. GTK and Qt apps expose a tree; Chromium, Electron and Firefox expose one only when started with accessibility on (--force-renderer-accessibility for Chromium and Electron), and Wine games expose none. An app with no tree is an error saying so, never a guessed one.

A shell started before Hyprland restarted keeps a stale HYPRLAND_INSTANCE_SIGNATURE. deskpilot then talks to the one instance whose socket answers, and with several live instances it stops and names them.

key --hold SECONDS holds each key down that long, up to 10, for auto-repeat or a game's movement.

key and type without --window press keys into the focused window. --take-focus focuses the window first and leaves it focused, for an app that ignores keys while unfocused. Keys go to the window's client, not through the compositor, so they never trigger a Hyprland keybind: use hyprctl dispatch for those.

paste hands text to a window as a clipboard only that window sees, then presses its paste chord (--keys, ctrl+v by default; most terminals want ctrl+shift+v). It pastes any text, including characters the keymap cannot type, and the user's clipboard stays as it is. XWayland windows share one clipboard, so paste refuses them: use type. clipboard reads the user's clipboard, and clipboard --set replaces it.

type sends keys as fast as the app reads them, about 1 KB a second into Chrome. Games read keys once per frame, so type into them with --delay 0.1 or more: each key is held that long, with the same gap before the next. For long text, paste is faster still.

wait-for waits for a window to appear, or with --gone to close. --focused, --title TEXT and --size WxH also wait for it to have keyboard focus, a title containing TEXT or that size. --changed SHOT waits until the window looks different from one of its shots, in all of it or in --region X Y W H of that shot's pixels: any pixel difference counts, and it prints a shot of the change with the box that changed.

Clicks, scrolls and keys from several agents at once run one after another in the order they arrive, so one type is never interleaved with another's keys. Up to 32 wait behind the running one; past that an action is refused.

MCP

deskpilot mcp serves the same actions over stdio: windows, focus, shot (returns the image and its shot id), click, move, scroll, hover (returns the image), drag, elements, key, type_text, paste_text, clipboard, copy_text, wait_for and wait_changed. click, move and scroll take shot, show and take_pointer; click, move, scroll and hover take element; shot, click, scroll, hover, drag, elements and wait_changed take hidden; click, key, type_text and paste_text take take_focus; key takes hold. It keeps one input device open for the session.

Every expected failure has a code, one stable word to branch on: not_found, ambiguous, not_visible, offscreen, stale, covered (another window is over the point), gone (the window closed), refused, busy, no_tree, unavailable, timeout or invalid. The CLI prints deskpilot: CODE: message. An MCP tool returns an error result whose text is CODE: message and whose structured content is {"code": ..., "message": ...}.

claude mcp add deskpilot -- deskpilot mcp

Limits

• type can only send characters the active keymap types; a character it lacks (such as é on a US layout) is an error rather than skipped. paste takes any text, except into XWayland windows.
• Hyprland plugins load only into the exact Hyprland they were built against, so deskpilot plugin load builds one per Hyprland commit and plugin source under ~/.local/share/deskpilot, and replaces a loaded build of other source.
• Each CLI action that uses the input device opens a new one and waits about a second for libinput to add it. The MCP server avoids that wait.
• drag stays within one window: a drag and drop into another app is not supported.
• Hyprland only, Linux only.

Development

AGENTS.md has the checks, the layout and the rules a change must keep, and CONTRIBUTING.md how to report a bug or send a change.