---

agent-device

CLI to control iOS and Android devices for AI agents influenced by Vercel’s agent-browser.

The project is in early development and considered experimental. Pull requests are welcome!

Features

• Platforms: iOS (simulator + limited device support) and Android (emulator + device).
• Core commands: open, back, home, app-switcher, press, long-press, focus, type, fill, scroll, scrollintoview, wait, alert, screenshot, close.
• Inspection commands: snapshot (accessibility tree).
• Device tooling: adb (Android), simctl/devicectl (iOS via Xcode).
• Minimal dependencies; TypeScript executed directly on Node 22+ (no build step).

Install

npm install -g agent-device

Or use it without installing:

npx agent-device open SampleApp

Quick Start

Use refs for agent-driven exploration and normal automation flows.

agent-device open Contacts --platform ios # creates session on iOS Simulator
agent-device snapshot
agent-device click @e5
agent-device fill @e6 "John"
agent-device fill @e7 "Doe"
agent-device click @e3
agent-device close

CLI Usage

agent-device <command> [args] [--json]

Basic flow:

agent-device open SampleApp
agent-device snapshot
agent-device click @e7
agent-device fill @e8 "hello"
agent-device close SampleApp

Debug flow:

agent-device trace start
agent-device snapshot -s "Sample App"
agent-device find label "Wi-Fi" click
agent-device trace stop ./trace.log

Coordinates:

• All coordinate-based commands (press, long-press, focus, fill) use device coordinates with origin at top-left.
• X increases to the right, Y increases downward.

Command Index

• open, close, home, back, app-switcher
• snapshot, find, get
• click, focus, type, fill, press, long-press, scroll, scrollintoview, is
• alert, wait, screenshot
• trace start, trace stop
• settings wifi|airplane|location on|off
• appstate, apps, devices, session list

Backends (iOS snapshots)

Backend	Speed	Accuracy	Requirements
xctest	Fast	High	No Accessibility permission required
ax	Fast	Medium	Accessibility permission for the terminal app, not recommended

Notes:

• Default backend is xctest on iOS.
• Scope snapshots with -s "<label>" or -s @ref.
• If XCTest returns 0 nodes (e.g., foreground app changed), agent-device falls back to AX when available.

Flags:

• --version, -V print version and exit
• --platform ios|android
• --device <name>
• --udid <udid> (iOS)
• --serial <serial> (Android)
• --activity <component> (Android; package/Activity or package/.Activity)
• --session <name>
• --verbose for daemon and runner logs
• --json for structured output
• --backend ax|xctest (snapshot only; defaults to xctest on iOS)

Skills

Install the automation skills listed in SKILL.md.

npx skills add https://github.com/callstackincubator/agent-device --skill agent-device

Sessions:

• open starts a session. Without args boots/activates the target device/simulator without launching an app.
• All interaction commands require an open session.
• If a session is already open, open <app> switches the active app and updates the session app bundle.
• close stops the session and releases device resources. Pass an app to close it explicitly, or omit to just close the session.
• Use --session <name> to manage multiple sessions.
• Session scripts are written to ~/.agent-device/sessions/<session>-<timestamp>.ad when recording is enabled with --save-script.
• Deterministic replay is .ad-based; use replay --update (-u) to update selector drift and rewrite the replay file in place.

Find (semantic):

• find <text> <action> [value] finds by any text (label/value/identifier) using a scoped snapshot.
• find text|label|value|role|id <value> <action> [value] for specific locators.
• Actions: click (default), fill, type, focus, get text, get attrs, wait [timeout], exists.

Assertions:

• is predicates: visible, hidden, exists, editable, selected, text.
• is text uses exact equality.

Replay update:

• replay <path> runs deterministic replay from .ad scripts.
• replay -u <path> attempts selector updates on failures and atomically rewrites the same file.
• Refs are the default/core mechanism for interactive agent flows.
• Update targets: click, fill, get, is, wait.
• Selector matching is a replay-update internal: replay parses .ad lines into actions, tries them, snapshots on failure, resolves a better selector, then rewrites that failing line.

Update examples:

# Before (stale selector)
click "id=\"old_continue\" || label=\"Continue\""

# After replay -u (rewritten in place)
click "id=\"auth_continue\" || label=\"Continue\""

# Before (ref-based action from discovery)
snapshot -i -c -s "Continue"
click @e13 "Continue"

# After replay -u (upgraded to selector-based action)
snapshot -i -c -s "Continue"
click "id=\"auth_continue\" || label=\"Continue\""

Android fill reliability:

• fill clears the current value, then enters text.
• type enters text into the focused field without clearing.
• fill now verifies the entered value on Android.
• If value does not match, agent-device clears the field and retries once with slower typing.
• This reduces IME-related character swaps on long strings (e.g. emails and IDs).

Settings helpers (simulators):

• settings wifi on|off
• settings airplane on|off
• settings location on|off (iOS uses per-app permission for the current session app) Note: iOS wifi/airplane toggles status bar indicators, not actual network state. Airplane off clears status bar overrides.

App state:

• appstate shows the foreground app/activity (Android). On iOS it uses the current session app when available, otherwise it falls back to a snapshot-based guess (AX first, XCTest if AX can’t identify).
• apps --metadata returns app list with minimal metadata.

Debug

• agent-device trace start
• agent-device trace stop ./trace.log
• The trace log includes snapshot logs and XCTest runner logs for the session.
• Built-in retries cover transient runner connection failures, AX snapshot hiccups, and Android UI dumps.
• For snapshot issues (missing elements), compare with --raw flag for unaltered output and scope with -s "<label>".

Boot diagnostics:

• Boot failures include normalized reason codes in error.details.reason (JSON mode) and verbose logs.
• Reason codes: BOOT_TIMEOUT, DEVICE_UNAVAILABLE, DEVICE_OFFLINE, PERMISSION_DENIED, TOOL_MISSING, BOOT_COMMAND_FAILED, UNKNOWN.
• Android boot waits fail fast for permission/tooling issues and do not always collapse into timeout errors.

App resolution

• Bundle/package identifiers are accepted directly (e.g., com.apple.Preferences).
• Human-readable names are resolved when possible (e.g., Settings).
• Built-in aliases include Settings for both platforms.

iOS notes

• Input commands (press, type, scroll, etc.) are supported only on simulators in v1 and use the XCTest runner.
• alert and scrollintoview use the XCTest runner and are simulator-only in v1.
• Real device support (including snapshots) is on the roadmap for iOS.

Testing

pnpm test

Useful local checks:

pnpm typecheck
pnpm test:unit
pnpm test:smoke

Build

pnpm build

Environment selectors:

• ANDROID_DEVICE=Pixel_9_Pro_XL or ANDROID_SERIAL=emulator-5554
• IOS_DEVICE="iPhone 17 Pro" or IOS_UDID=<udid>
• AGENT_DEVICE_IOS_BOOT_TIMEOUT_MS=<ms> to adjust iOS simulator boot timeout (default: 120000, minimum: 5000).

Test screenshots are written to:

• test/screenshots/android-settings.png
• test/screenshots/ios-settings.png

Contributing

See CONTRIBUTING.md.

Made at Callstack

agent-device is an open source project and will always remain free to use. Callstack is a group of React and React Native geeks. Contact us at hello@callstack.com if you need any help with these technologies or just want to say hi.