computer-use

Cross-platform desktop automation for agents: screenshots, mouse, keyboard, scroll, displays, accessibility trees, OCR, CDP attach, isolation, and a permission doctor. One Rust binary on macOS, Linux, and Windows. The same actions are available as a Rust library.

The smart-computer-use agent skill shells out to this binary.

Install

GitHub Releases

Download a build for your platform from GitHub Releases:

Asset	Platform	Notes
`computer-use-linux-x64.tar.gz`	Linux x86_64	Built with `--features portal` (ScreenCast + RemoteDesktop + PipeWire).
`computer-use-macos-arm64.tar.gz`	macOS Apple silicon	Default features. `Computer Use.app` is bundled as a separate zip when the helper is built (see `docs/PACKAGING.md`).
`computer-use-windows-x64.zip`	Windows x86_64	Default features.

Each archive contains the computer-use binary (computer-use.exe on Windows). A matching <asset>.sha256 file and a combined SHA256SUMS file are attached to the release.

# macOS (Apple silicon)
curl -fsSL -o computer-use-macos-arm64.tar.gz \
  https://github.com/lindoelio/computer-use/releases/latest/download/computer-use-macos-arm64.tar.gz
tar -xzf computer-use-macos-arm64.tar.gz
sudo install -m 755 computer-use /usr/local/bin/computer-use

# Linux x86_64
curl -fsSL -o computer-use-linux-x64.tar.gz \
  https://github.com/lindoelio/computer-use/releases/latest/download/computer-use-linux-x64.tar.gz
tar -xzf computer-use-linux-x64.tar.gz
sudo install -m 755 computer-use /usr/local/bin/computer-use

Check the archive before you install it:

shasum -a 256 -c computer-use-macos-arm64.tar.gz.sha256
# or, on Linux:
sha256sum -c computer-use-linux-x64.tar.gz.sha256

cargo

cargo install computer-use

Rust 1.91 is the supported minimum. On Linux, install the native libraries listed under Permissions before building.

Usage

Every action writes one JSON object to stdout and a trailing newline. A failed action exits with status 1 and a {"ok":false,"error":{...}} body. A bad command line (missing flags, unknown arguments) exits with status 2 and still prints JSON. --help and --version stay human-readable and exit 0.

computer-use --help
computer-use --version
computer-use wait --ms 10
computer-use --pretty wait --ms 10

serve reads newline-delimited JSON from stdin and writes one JSON response per line. --json does the same. exec runs one JSON action from an argument and honors --pretty. There is no TCP server in any release.

printf '%s\n' '{"id":1,"action":"wait","ms":10}' | computer-use serve
printf '%s\n' '{"id":1,"action":"click","x":10,"y":20}' | computer-use --json
computer-use exec '{"action":"wait","ms":10}'
computer-use --pretty exec '{"id":"a","action":"displays"}'

--adapter native|openai|anthropic rewrites the exec and stdin requests on the way in. The native protocol is what Action and the action-name list describe below; the adapter layers translate the OpenAI computer and Anthropic computer_20251124 shapes into the native one and set coord=image. See docs/CLI.md for the full JSON schema and --help for the surface area.

Default max-width

screenshot (and zoom) scale the image down so its width is at most 1440 pixels by default. Pass --max-width 0 to keep the source width. --max-width N with any other positive integer caps the width at N. --canvas WxH letterboxes the result into WxH without scaling up. On a Retina display the image pixel size can be scale times the logical size.

Policy path

~/.config/computer-use/policy.toml on every OS, including macOS and Windows. The admin file is /etc/computer-use/policy.toml on Linux and macOS, and %ProgramData%\computer-use\policy.toml on Windows. --policy <PATH> overrides the user file. A missing file is default allow plus the built-in deny list. --policy check --app X reports what the effective policy would decide without performing the action.

Feature flags

Flag	What it pulls in	Default on this binary
(default)	Screenshots, mouse, keyboard, accessibility tree, OCR, isolation, the macOS helper `.app` plumbing.	always
`cdp`	`chromiumoxide` 0.9.1 + a small `tokio`. Required for `computer-use --cdp browser ...`.	off
`portal`	`ashpd` 0.13.13 ScreenCast + RemoteDesktop, `pipewire` 0.9, `reis` 0.7. Linux only.	on in the Linux release binary; off in the default CI build
`skylight`	macOS private SkyLight focus path behind `--tier synthetic` / `--synthetic-focus`.	off
`virtual-display`	macOS private `CGVirtualDisplayDescriptor` probe; probe-only, never creates a display.	off

--features cdp,--features portal,--features skylight,--features virtual-display build what the flag page describes. --all-features compiles everywhere (Linux builds also link PipeWire).

Top-level commands

screenshot          Capture the screen, a region, or one window, and write a PNG or JPEG file
click               Move the pointer to a logical point and click, or press an element
triple-click        Click three times. Alias of `click --count 3`
press               Press an element from the current tree. `--count` may be 1, 2, or 3
set-value           Set an element's value. Secure fields are refused
select-text         Select a character range. Secure fields are refused
perform-action      Perform an allowlisted accessibility action
scroll-to           Scroll an element into view
focus               Focus an element
move                Move the pointer without clicking
drag                Drag from the first point through the rest, then release
mouse-down          Press a mouse button and leave it down until `mouse-up`
mouse-up            Release a mouse button pressed by `mouse-down`
type                Type Unicode text into the focused control
key                 Press a key combination
hold-key            Hold one key combination, then release it
clipboard           Read or write the clipboard (`get`, `set`)
zoom                Screenshot one region. With `--app`/`--window`, the region is window-local
scroll              Scroll the mouse wheel
get-state           Report displays, cursor, frontmost window, and a best-effort accessibility tree
wait                Sleep, then return `waited_ms`
displays            List connected displays without capturing pixels
isolation           Report what this build can isolate and what it cannot
list-apps           List running applications
list-windows        List on-screen windows in z-order. Index 0 is front
launch              Launch an application by name, bundle id, path, or pid
activate            Bring an application forward
quit                Ask an application to quit
policy              Show whether policy allows an app. This does not perform an action
pause               Pause mutating actions until `resume`
resume              Clear the pause flag
doctor              Report permissions, the TCC identity, and how to fix what is missing
tree                Print the key window's accessibility tree
ocr                 Run OCR on the most recent capture for `--app` or `--window`
menu                Walk a running app's menu bar (macOS only)
apple-event         Send an Apple event script via `osascript` (macOS only)
shortcuts           Run a Shortcuts workflow by name (macOS only)
dock                List or click Dock items (macOS only)
spaces              Group on-screen windows by the screen that contains each window's center
session             Session cache and accessibility teardown (`end`, `prune`)
serve               Read newline-delimited JSON actions from stdin
exec                Run one JSON action from an argument
overlay             Local agent-cursor helper. This is not a network service
browser             Chrome DevTools Protocol (CDP) subcommands. Requires `--cdp`
help                Print this message or the help of the given subcommand(s)

Permissions

Input and capture need permission from the operating system. get-state still returns ok: true when a subsystem is denied: the missing field is null and note says why. doctor enumerates each required grant plus the TCC identity; the JSON result is the source of truth.

macOS

Permission	What uses it	Where to grant
Screen Recording	`screenshot`, `zoom`, the `capture` path in `get-state`, ScreenCaptureKit window capture	System Settings → Privacy & Security → Screen Recording
Accessibility	The accessibility tree, hit-testing, `AXPress`, `set_value`, `select_text`, element `type`, `menu`, Apple events, `osascript` fallback	System Settings → Privacy & Security → Accessibility
Post Event	Tier `event` (`CGEventPostToPid`), `CGEventCreate`, hit-test	System Settings → Privacy & Security → Accessibility
Input Monitoring	The P6 kill-switch tap	System Settings → Privacy & Security → Input Monitoring

The bundle id is ai.lindoelio.computer-use when the binary is run from Computer Use.app (scripts/package-macos-app.sh builds it). A bare binary does not have a stable bundle id — the TCC grant is for the signed bundle. The --prompt flag on doctor calls the macOS request APIs without changing anything in this process otherwise; --open-settings opens the first missing pane.

Windows

Run the process in a session that can send input (an interactive desktop, not Session 0).

Apps and windows come from EnumWindows, GetWindowTextW, GetWindowThreadProcessId, and QueryFullProcessImageNameW (windows 0.62). There is no PowerShell anywhere in the binary.
The accessibility tree is the UI Automation control-view walker (uiautomation 0.25.1). tree reports source: "uia", --app pid:NNN is a deterministic selector, and GetForegroundWindow is never used to resolve a window the caller already named.
Element actions run through UIA control patterns (Invoke, Value, Text, ScrollItem, ExpandCollapse) in the background — tier ax, no cursor motion. IsPassword controls refuse set_value and type --element. When a pattern is missing, or its single call fails (some Win32 apps ignore Invoke until they are in front), the action is retried as tier activate: SetForegroundWindow plus enigo input, under the same foreground.lock as every other foreground path, with a foreground required warning on the result. --require-background refuses that fallback, and the P6 policy allowlist still applies.
Per-window capture is one Windows.Graphics.Capture frame from the resolved HWND (windows-capture 2.0.1). If graphics capture cannot produce a frame, the desktop image is cropped instead and reported as capture_source: "xcap", so covered pixels may be wrong.
OCR is Windows.Media.Ocr.OcrEngine for the user-profile languages. If no recognizer language pack is installed, ocr returns execution_failed naming Settings → Apps → Optional features.
doctor reports accessibility (did UIAutomation::new succeed), plus best-effort uiaccess and integrity read from the process token with OpenProcessToken / GetTokenInformation. Neither is a required grant.

A corporate policy that blocks UI Automation or simulated input surfaces as execution_failed or a null field.

Linux

The binary talks to X11 through libxdo (enigo) and captures through xcap (XCB, RandR, D-Bus). On a Wayland session, the portal feature is needed for capture (ScreenCast) and event-tier input (RemoteDesktop). Install the build packages on Debian/Ubuntu:

sudo apt-get install -y pkg-config libxdo-dev libx11-dev libxcb1-dev \
  libxrandr-dev libdbus-1-dev libxkbcommon-dev libatspi2.0-dev \
  libpipewire-0.3-dev libclang-dev clang

xdotool on PATH improves the frontmost-window lookup. The accessibility tree is AT-SPI (default features). Tier event is X11 XTest on plain X11 and the RemoteDesktop portal on Wayland when the portal feature is compiled in. Without portal, Wayland tier event is unsupported. Grant the session access to the X display (DISPLAY set, and the same user that owns the session).

The release binary is built with --features portal, which links libpipewire-0.3 at build time and pulls one frame off the portal's PipeWire stream. Linux OCR is unsupported in 0.2.0.

Input tiers

--tier selects the input path. Auto (the default) is platform-tuned:

macOS: ax for presses and value changes; event then activate for coordinates, type without --element, key, and scroll. synthetic is never automatic.
Windows: ax (UIA control patterns); falls back to activate (SetForegroundWindow + enigo) under foreground.lock when a pattern is missing or fails. Auto is activate for free-text type, key, and scroll.
Linux: ax (AT-SPI) when present; event (XTest on X11, RemoteDesktop portal on Wayland) for coordinates; activate (enigo) as the foreground fallback.

`--tier`	What it does
`ax`	Accessibility action or value. Does not activate the app. Also the scroll-bar path for `scroll`.
`event`	Post events to the target process. Does not move the system cursor. macOS posts to the target pid via `CGEventPostToPid`. Linux uses XTest on X11 and the RemoteDesktop portal on Wayland.
`synthetic`	Private SkyLight front-process call, then post events. Requires `--features skylight` and `--tier synthetic` (or `--synthetic-focus`).
`activate`	Activate the target, run foreground input, then restore the previous app. Takes `foreground.lock`.

--require-background refuses activate and turns cursor motion into execution_failed. --allow-foreground skips the lock. --no-overlay skips the agent-cursor helper socket. The agent cursor overlay (overlay serve) is macOS-only; a missing helper adds overlay helper not running and the action still succeeds.

Scroll is verified on macOS, not assumed

A scroll wheel posted with CGEventPostToPid is not delivered to the views of a non-frontmost app: the post returns without error and nothing moves. So scroll on macOS measures instead of assuming. It reads the AXScrollArea scroll bar (under --x/--y, else the window's first usable scroll area) before and after the post, then reports what actually happened:

`via`	When
`event`	The posted wheel event moved the bar.
`ax_scrollbar`	The wheel event did nothing, so `AXValue` was written on the bar. Adds the warning `wheel event not delivered to background app; used AX scroll bar`. Neither path activates the app or moves the cursor.

scrolled is true only when a bar value really changed, and before / after carry the values (0.0–1.0) on the primary axis. A bar already at its limit returns ok with scrolled: false and a warning naming the edge; a scroll that cannot be verified at all (no AXScrollArea) is execution_failed, never a silent success. --tier ax scrolls by driving the bar directly and posts no event.

Two honest limitations:

The wheel event is not delivered even to a frontmost app. Measured live on macOS 27 with TextEdit frontmost: the posted wheel left AXValue untouched, so the fallback ran. The verification is what makes that visible instead of silent. The warning text still says "background app" because that contract is fixed.
The step is an approximation. AX does not say how much of a document is on screen, and an AXScrollBar element spans the whole track, so the derived fraction sits near 1.0 and one wheel line moves roughly a third of the content. That is close to the system default of three lines per notch, but it is a coarse scroll, not a precise one. Pass an explicit --amount for control.

Windows and Linux do not verify this way; their scroll responses keep their existing shape.

JSON actions

The Action enum in src/protocol.rs is the protocol. id is echoed on the response. Unknown fields that belong to the action are rejected. Global flags (app, window, element, diff, full, return_state, include_screenshot, coord, session, tier, adapter, display, require_visible, policy, confirm_cmd, allow_foreground, require_background, no_overlay, cdp, cdp_url) are siblings of action and stripped before the action is decoded.

`action`	Fields	Notes
`screenshot`	`output`/`path`, `region` (`"x,y,w,h"` or `{x,y,width,height}`), `format` (`png`/`jpeg`/`jpg`), `max_width`, `canvas` (`"WxH"`), `base64`, `marks`	Default `max_width` is 1440. `0` keeps the source width. `marks` stamps tree indices (`n`) on a window target or OCR indices (`oN`) when the tree is empty.
`click`	`x`, `y`, `button` (`left`/`right`/`middle`/`back`/`forward`), `double`, `count`	`count` is 1, 2, or 3. `double: true` means `count: 2` when omitted. With `--element`/`element` the screen coordinates are ignored.
`double_click`	`x`, `y`, `button`	Alias: `click --double`.
`triple_click`	`x`, `y`, `button`	Alias: `click --count 3`.
`move`	`x`, `y`	Alias: `move_mouse`.
`press`	`count` (1–3), `button`	AXPress for left, AXShowMenu for right when advertised.
`type`	`text`	Uses `AXValue` when `--element` is set and settable.
`key`	`keys`	Combo such as `cmd+shift+4`.
`scroll`	`x`, `y`, `dx`/`scroll_x`, `dy`/`scroll_y`, `direction`, `amount`	Positive `dy` is down; positive `dx` is right. `direction` + `amount` is a separate form; do not mix. macOS adds `via`, `scrolled`, `before`, `after` (see Scroll is verified on macOS).
`drag`	`path` (array of points or `x,y,x,y` string), `button`	At least two points. Tier `event` posts mouse-dragged events to the pid; tier `activate` uses enigo.
`mouse_down`	`x`, `y`, `button`	Held until `mouse-up` or process exit.
`mouse_up`	`x`, `y`, `button`	Releases the press from `mouse_down`.
`hold_key`	`keys`, `ms` (≤ 30000)	One combination. Space-separated sequences are refused.
`zoom`	`region` (required), `output`/`path`, `format`, `max_width`, `canvas`, `base64`	Window-local coordinates when `--app`/`--window` is set.
`clipboard_get`	none	Text wins; `kind: "empty"` is reported. An image is written to the session cache and `kind: "image"` is reported with a `path`.
`clipboard_set`	`text`, `image` (PNG path)	Requires a confirm hook. Empty `text` is `invalid_arguments`.
`focus`	none	Sets `AXFocused` on the `--element`.
`set_value`	`value`	Refused for secure fields (`policy_denied`; the value is never read).
`select_text`	`start`, `end`	Refused for secure fields.
`perform_action`	`name`	Allowlist (`AXPress`, `AXShowMenu`, `AXIncrement`, `AXDecrement`, `AXConfirm`, `AXCancel`, `AXRaise`, `AXScrollToVisible`, the four page-scroll actions).
`scroll_to`	none	Aliases: `scroll-to`, `scroll_to_visible`, `scroll-to-visible`.
`get_state`	none	Alias: `get-state`.
`cursor_position`	none	Reads the pointer. The Anthropic adapter maps `cursor_position` here.
`wait`	`ms`, `seconds`	Summed when both are set. Fractions are allowed in `seconds`.
`displays`	none	List displays without capturing pixels.
`isolation`	none	Pure report: what this build can isolate.
`list_apps`	`all`	`all: true` includes accessory and prohibited apps.
`list_windows`	none	z-order; index 0 is front.
`launch`	`app` (sibling)	Launch via `NSWorkspace` on macOS; otherwise `unsupported`.
`activate`	`app`, `window`	`--window` is checked first.
`quit`	`force`	`--force` is a confirm-gated terminate.
`policy_check`	`app` (sibling)	Reports `allow`/`deny` without performing the action.
`pause` / `resume`	none	Cross-process kill switch. macOS only (CGEventTap).
`doctor`	`prompt` (alias `request`), `open_settings`, `x11`	Envelope `ok: true` even when a permission is missing.
`tree`	none	Numbered accessibility tree for the key window.
`ocr`	`force`	Re-run OCR even when `--app` already produced it.
`session_end`	none	Restore Electron accessibility flags this process changed.
`session_prune`	`dry_run`, `ttl_secs`	Remove empty session directories and stale ones. Default TTL is 24h.
`menu_list`	`app` (override), `depth`	macOS only. Walks `AXMenuBar → AXMenuBarItem → AXMenu`.
`menu_click`	`app` (override), `path` (`"File > Save"`)	macOS only. `AXPress`es the leaf.
`apple_event`	`app` (override), `script`, `language` (`applescript`/`javascript`)	macOS only. Runs `osascript`. The script body is never logged.
`shortcuts_run`	`name`	macOS only. Invokes `shortcuts run <name>`.
`dock_list`	none	macOS only. Lists Dock items.
`dock_click`	`title`	macOS only. Case-insensitive title match.
`spaces`	none	Groups on-screen windows by screen. Mission Control moves are not attempted.
`batch`	`actions`, `stop_on_error`	One resolved target, one revision sequence. Each step re-validates. Nested batch is `invalid_arguments`.
`browser_tabs`	none	`cdp` feature.
`browser_activate`	`target_id`	`cdp` feature.
`browser_navigate`	`url`, `target_id`	`cdp` feature. URL is checked against `[browser].blocked_urls`.
`browser_click`	`selector`, `target_id`	`cdp` feature. CSS selector, clicked via `Input.dispatchMouseEvent`.
`browser_text`	`selector`, `target_id`	`cdp` feature. Returns the matched element's `innerText`.
`browser_type`	`text`, `selector` (optional), `target_id`	`cdp` feature. Selector or page body.
`browser_eval`	`script`, `target_id`	`cdp` feature. The script body is never logged verbatim; only its length and a stable `DefaultHasher` hash.

{"id":1,"action":"click","x":10,"y":20,"button":"left"}
{"id":2,"action":"key","keys":"cmd+shift+4"}
{"id":3,"action":"scroll","direction":"down","amount":3}
{"id":4,"action":"screenshot","region":"0,0,800,600","format":"png"}
{"action":"get_state"}

The --adapter openai and --adapter anthropic translations cover click, double_click, scroll, type, keypress, move, wait, screenshot, cursor_position, back, forward, drag, mouse_down, mouse_up, triple_click, left_click_drag, hold_key, and zoom.

Error codes

The JSON error.code is one of the strings below. ComputerError::code() is the stable contract.

Code	When
`invalid_arguments`	Bad CLI flags or malformed JSON.
`execution_failed`	Framework error that is not in this table.
`io_error`	stdin/stdout/filesystem failure.
`unsupported`	Off-platform or feature-off.
`app_not_found`	Name, bundle id, or pid matched nothing running, and launch was not requested.
`window_not_found`	App resolved, window did not.
`element_not_found`	Index missing in the cached or fresh tree, or there is no session cache.
`stale_element`	The path's role, identifier, or frame (beyond 2 points) changed, or the cached pid, window, or revision header does not match. Action not performed. Body includes the new tree.
`permission_required`	Accessibility or Screen Recording is off. The `doctor` `fix` object names the Settings pane.
`policy_denied`	Built-in or configured deny.
`confirm_required`	Sensitive action and no confirm hook is configured.
`confirm_denied`	Hook exited non-zero or timed out.
`blocked_url`	Browser URL matched the block list.
`paused`	Local user input tripped the kill switch.
`cdp_unavailable`	Feature off, or Chrome has no debugging port.
`display_unavailable`	The screen is locked, the display is asleep, or no display or shareable window is available. Checked before `window_not_found`; no xcap crop fallback.

Library

use computer_use::{dispatch, Action};

let response = dispatch(Action::Wait {
    ms: Some(0),
    seconds: None,
});
assert!(response.ok);

DesktopBackend (also exported as Backend) is the platform trait. create_desktop_backend() returns the backend for the current operating system. dispatch and dispatch_with share the path used by the CLI. Action and the JSON request parser use serde with tag = "action". Errors use thiserror (ComputerError). The full surface (CallContext, LineOptions, handle_request_line_opts) is documented in docs/CLI.md and lives in src/protocol.rs.

Coordinates

Click, move, scroll, and screenshot regions use logical points. A display's width, height, x, and y are logical. scale is physical pixels divided by logical width (often 2.0 on Retina). A screenshot's width and height are image pixels, which can be larger than the logical region. The map lives at ${XDG_CACHE_HOME:-~/.cache}/computer-use/sessions/<session>/screenshot-map.json (macOS and Linux) or %LOCALAPPDATA%\computer-use\cache\sessions\<session>\screenshot-map.json (Windows), mode 0600.

--coord image (or coord=image) interprets x/y as image pixels of the most recent screenshot; the map remaps them to screen points. The OpenAI and Anthropic adapters set coord=image automatically and require a stored map.

--display <id> resolves a monitor from displays and gates every screen point, --region, and --window against it. Points outside the monitor are invalid_arguments. The filtered capture reuses the shared backend::capture path, so the map origin is the monitor origin. --display virtual is Linux-only; it prints the Xvfb recipe and never captures.

Isolation

isolation is a pure report: {platform, display_filter, cdp, virtual_display, set_of_marks}. macOS reports virtual_display: "experimental" only when both CGVirtualDisplayDescriptorCreate and CGVirtualDisplayCreate resolved; otherwise unsupported. The build never creates a virtual display — see docs/ISOLATION.md for the reason and the VM proxy design.

Development

cargo fmt --all
cargo build
cargo test --all-targets
cargo clippy --all-targets --all-features -- -D warnings

Tests do not move the pointer, capture the screen, or require Screen Recording or Accessibility. macOS-only tests behind #[cfg(target_os = "macos")] are skipped on Linux and Windows runners. The CI matrix is Ubuntu (default + --features portal), macOS (default, with bash -n/shellcheck on scripts/package-macos-app.sh and an ad-hoc bundle check on the debug binary), and Windows (default).

~/.grok-prompts/linux-check.sh runs the same checks inside rust:1-bookworm when Docker is available; otherwise CI is the gate.

Documentation

docs/CLI.md — every command, flag, JSON action, and error code, generated from --help and the protocol enums.
docs/ROADMAP.md — the phase plan, the progress table, and the parity audit.
docs/ISOLATION.md — display filter, Xvfb recipe, virtual-display limitations.
docs/PACKAGING.md — macOS .app packaging, signing, notarization.
docs/SKYLIGHT.md — synthetic-focus design notes.
docs/policy.example.toml — policy file reference.

License

Licensed under either of

Apache License, Version 2.0 (LICENSE-APACHE)
MIT license (LICENSE-MIT)

at your option.

Notes from live testing

Element indices across CLI calls. Indices from tree are cached per session. The default session is the current process, so separate one-shot invocations only share indices when you pass the same --session <ID> to tree and to the element action.
set_value replaces the field. set-value (and type --element N without an explicit --tier) writes the AX value, replacing the whole field contents. Pass --tier event to insert keystrokes at the focused caret instead.
type --tier event needs keyboard focus. Keystrokes posted to a pid are dropped if no element in that app has focus. type then returns ok with a warning; click the field (click --tier event) or pass --element N first.
scroll --tier event used to lie. It returned ok without scrolling a background app, because a wheel event posted to a non-frontmost pid is never delivered to its views. scroll now reads the scroll bar back and reports via, scrolled, before, and after; when the wheel event is not delivered it falls back to the AXValue scroll bar and says so in a warning. If there is no scroll area to verify against, it is execution_failed instead of a silent success.

computer-use 0.2.0