Platform support

Proven operating systems, Linux window systems, and browser routes, with known boundaries and remaining work.

A capability counts as supported only when a Rust harness test observes the result in application- or desktop-owned state (see How Cua Driver works). Unsupported routes return a structured refusal.

Level	Meaning
Supported	Proven against app-owned or desktop-owned state.
Supported with limits	Common paths proven; some delivery shapes refuse.
Experimental	Backend exists; coverage incomplete. Assume nothing unlisted works.

Operating systems#

Platform	APIs	State
Windows	Win32, UI Automation, native input, window messages	Supported. Coverage: Electron, Tauri, WPF, WinUI 3, WebView2. Some background Chromium gestures and elevated-integrity targets refuse or are unproven.
macOS	AppKit, Accessibility, Quartz/HID, ScreenCaptureKit	Supported. Coverage: Electron, Tauri, AppKit, SwiftUI, WKWebView. Needs Accessibility and Screen Recording. Some background scroll and drag shapes refuse.
Linux X11	X11/EWMH, XTest, AT-SPI	Supported with toolkit limits. Toolkits that reject synthetic background events get an explicit refusal.
Linux Wayland	AT-SPI plus compositor-specific discovery, capture, activation, and portal input	Supported with compositor limits. Semantic background actions work where the app exposes them; raw input to an occluded window generally does not.

Linux window systems#

"Wayland" is not one API: each compositor exposes different discovery, capture, activation, and input. Native Wayland is opt-in with CUA_DRIVER_RS_ENABLE_WAYLAND=1; without it, Wayland sessions use XWayland routes where the app has a real X11 window.

Environment	State	Proven	Main limits
X11/Xorg	Supported	Discovery, capture, AT-SPI, foreground input, semantic background actions, desktop scope	Raw background delivery depends on the toolkit.
Sway (wlroots)	Supported with limits	The full Electron, Tauri, GTK, and capture catalog	Focus-bound raw background input refuses. Other wlroots compositors unproven.
GNOME/Mutter	Supported with limits	AT-SPI, GTK controls, compositor geometry and capture, foreground activation, portal/libei input	Needs the bundled WinRects Shell helper and one Shell restart. Portal video recording incomplete.
KDE/KWin	Experimental	Plasma 6 startup, GTK AT-SPI discovery, portal availability	The KWin helper is read-only; raw target-addressed input refuses.
Hyprland / Omarchy	Experimental	Qualified isolated input since 0.24.0	Does not inherit Sway coverage.
`cua-compositor` (nested)	Experimental	Native GTK behavior, capture, per-surface raw input	Owns its own compositor; says nothing about stock Wayland.
XWayland	Supported with limits	X11 routes for real X11 windows; AT-SPI fallback	Depends on whether the app actually uses X11.

X11 window capture supports packed 16-, 24-, and 32-bit TrueColor and DirectColor images. DirectColor capture reads the window's current colormap for each frame, including palette changes that do not repaint the window. Other indexed visual classes remain outside the native capture path.

Distribution names (Ubuntu, Arch) do not matter much; what matters is the display session and the app's backend. A portal grant belongs to the runtime that received it and may prompt again after a restart.

Hyprland and Omarchy#

Omarchy is an Arch setup built on Hyprland. Modern Hyprland is not wlroots, so Sway results do not carry over. An optional Hyprland plugin, off by default and outside the normal installer, adds two independent agent seats (Cua-Agent, Cua-Agent-2) for isolated background input. Cua Driver 0.24.0 shipped it with a narrow qualification:

Apps: Calc from libreoffice-fresh 26.2.5-3 and Inkscape 1.4.4-6, checked against package metadata before every action.
The plugin's Hyprland ABI and compiler toolchain must match the compositor.
Not covered: Chromium, Electron, and XWayland raw input; Unicode and IME; multiple layouts; modified pointer gestures.
Explicit foreground actions change focus and pointer and do not restore them.

Evidence for this feature is recorded in PRs #3547, #3557, and #3572. General Hyprland support still needs an accepted input and permission design, packaged install and upgrade proof, and the full native catalog on an integrated candidate.

Browser routes#

Browser mutation always starts from an exact native (pid, window_id) binding.

Surface	Proven	Limit
Chrome and Edge on Windows	Snapshot, navigate, type, trusted background click, DOM pointer actions, dialogs, file assignment, downloads, frames, multi-tab	Elevated-integrity targets refuse.
Chrome on macOS; Chrome and Edge on Linux X11	The same, except trusted pointer input	Trusted pointer returns `browser_input_trust_unavailable`; Linux dialogs need foreground.
Chrome on Sway	Exact binding when compositor identity and geometry agree	Generic GNOME or KDE Wayland is read-only.
Electron (Windows, macOS, X11, Sway)	Typed mutation while one native window maps to one page	A second page or window invalidates the route.
Brave, Vivaldi, Opera, Arc	Detected, not release-accepted	Use at your own risk.
Tauri, WKWebView, WebKitGTK, split-process WebView2, Safari, Firefox	Browser identity and a clean refusal	Use native accessibility and pixel actions.

Logged-in profile attachment is covered in Browsers.

Boundaries and remaining work#

These are operating-system boundaries, not bugs; the driver refuses instead of working around them:

Windows integrity levels: a lower-integrity process cannot inject input into a higher-integrity (elevated) app.
Standard Wayland: an ordinary client cannot send raw input to an unfocused, occluded window. AT-SPI actions still work in the background.
macOS: off-Space SwiftUI trees, keyboard commits to minimized windows, and apps that accept only HID-tap input. Workarounds are in Troubleshoot.

Off-Space windows on macOS#

macOS omits windows on other Spaces from an app's AXWindows list. For AppKit windows, the driver recovers the requested window by its exact CGWindowID, so get_window_state, element actions, and background pixel actions keep working while the window stays on its Space. The user's current Space, frontmost app, and keyboard focus do not change. Recovery probes the first 2,000 accessibility element ids of the app; a window beyond that limit reports ax_window_unresolved. The agent cursor does not paint while an action targets a window on another Space. Windows virtual desktops and Linux workspaces do not yet suppress the cursor for off-workspace targets.

Desktop captures leave out the agent cursor#

get_desktop_state keeps the driver's own agent cursor and session pill out of the PNG it returns. Every response reports the outcome in structuredContent.agent_overlay_capture: excluded with the native method, not_present when no overlay pixels were on screen, or not_excluded with a reason.

Platform	How the cursor is left out
macOS 14 or later	ScreenCaptureKit display filter; the `screencapture` fallback reports `not_excluded`.
Windows 10 2004 or later	`WDA_EXCLUDEFROMCAPTURE` for the length of the capture; older builds report `not_excluded`.
Linux X11	The overlay hides for the root grab (`method` ends in `+x11_save_under` without a compositor).
Linux Wayland	The compositor draws the cursor into the output; captures report `not_excluded`.

Open evidence gaps, with no dates promised: broader WPF and WinUI 3 gestures, WebView2 native input, a controlled background_uipi_blocked fixture, more AppKit actions and SwiftUI popovers, real-Xorg multi-pointer, GNOME and KDE renderer catalogs, and the full nested-compositor catalog. A row changes only when an unchanged catalog cell produces new app-owned evidence at an exact source commit.

Contributors: source layout, harnesses, and runner commands are in the repository's libs/cua-driver/README.md and libs/cua-driver/docs/test-harnesses-guide.md.