RELATED 01
OpenComputerUse
Desktop and mobile app control for agents, run in background sessions over MCP
20GitHub stars
16popularity score
agent-device
Lets a coding agent drive and verify mobile, TV, desktop and web apps on simulators, emulators and devices.
By Callstack TypeScript Local · stdio TestingDeveloper tools
101 popularity score ★ 5.3k ↓ 684k/wk
Connect your client
Local (npm)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add agent-device -- npx -y agent-device mcpClaude Desktopclaude_desktop_config.json
{
"mcpServers": {
"agent-device": {
"command": "npx",
"args": [
"-y",
"agent-device",
"mcp"
]
}
}
}Cursor~/.cursor/mcp.json
{
"mcpServers": {
"agent-device": {
"command": "npx",
"args": [
"-y",
"agent-device",
"mcp"
]
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"agent-device": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"agent-device",
"mcp"
]
}
}
}Codex~/.codex/config.toml
[mcp_servers.agent-device]
command = "npx"
args = ["-y", "agent-device", "mcp"]Gemini CLITerminal
gemini mcp add --scope user agent-device npx -- -y agent-device mcpSettings 1
AGENT_DEVICE_NO_UPDATE_NOTIFIER |
Optional: Set to 1 to turn off the npm update notice. |
|---|
What it does
agent-device, from Callstack, lets a coding agent inspect, control and verify a running app so it can check its own changes. It works with iOS, Android and HarmonyOS on simulators, emulators and physical devices, plus tvOS, Android TV, Amazon Vega OS TV, web, macOS and Linux desktop apps.
The agent reads accessibility snapshots of the screen instead of reasoning over screenshots alone, acts on elements by ref or selector (tap, fill, scroll, swipe, gestures, alerts), asserts state, and saves evidence: screenshots, video, logs, traces, network data, performance samples, crash details and React Native component trees. Working steps can be saved as .ad replay scripts for CI or exported as Maestro YAML. Sessions are scoped to the git worktree so parallel agents do not take over each other's simulators, and the same commands can drive devices on BrowserStack, AWS Device Farm and Limrun.
It is one tool with three entry points: a CLI, a typed Node.js API, and a stdio MCP server started with agent-device mcp, which exposes the installed commands as structured tools and an MCP-only help tool. Install it with npm install -g agent-device (Node.js 22.12 or newer; 24 for web). Platform tooling such as Xcode or ADB is needed for the targets you use.
Tools 37
open- Open an app or URL on a target device or platform and start a session.
close- Close the session.
snapshot- Capture an accessibility snapshot of the current screen with refs for elements.
diff- Show what changed in the snapshot since the last one.
screenshot- Save a screenshot as evidence.
press- Tap or press an element by ref or selector, or at coordinates.
click- Click an element (web and desktop targets).
fill- Type a value into a text field.
type- Type text.
hover- Hover over an element.
scroll- Scroll the screen.
swipe- Swipe on the screen.
gesture- Perform gestures such as pinch or multi-touch.
longpress- Long-press an element.
back- Go back.
home- Go to the home screen.
find- Find elements semantically, for example by text, and act or assert on them.
get- Read an element's text or attributes.
is- Assert an element's state, such as visible.
wait- Wait for text or an element to appear.
alert- Read or handle system alerts.
apps- List installed apps.
install- Install an app build in place.
reinstall- Reinstall an app for a fresh state.
devices- List available devices, simulators and emulators.
settings- Change device settings such as Wi-Fi.
clipboard- Read or write the device clipboard.
keyboard- Control the on-screen keyboard.
push- Simulate a push notification.
logs- Stream and read app logs for the session.
network- Dump network traffic seen by the app.
perf- Sample performance data such as frames and memory.
record- Record a video of the session.
trace- Record a trace.
replay- Run a saved .ad replay script.
batch- Run several commands in one call.
help- Return usage guides for commands and topics.