agent-device MCP Server: Setup for Claude Code, Cursor and Other Apps

• Opens an app on a phone, tablet, TV or desktop and shows its screen

  Starts a session and opens the named app on a simulator, emulator or device. It returns the first snapshot of tappable items, each with a reference like @e2.

  Try: “Open the Contacts app on the iOS simulator.”

  Runs code · open

• Reads what is on screen as a list of labelled items

  Reads the app's accessibility tree and returns buttons, fields and text with short references the agent can act on. This uses fewer tokens than a screenshot.

  Try: “What is on the screen right now?”

  Read · snapshot

• Takes a picture of the screen

  Saves a screenshot, useful as evidence or when the accessibility data is poor.

  Try: “Take a screenshot of the checkout screen.”

  Files · screenshot

• Taps a button or item

  Presses an item by its reference, label or coordinates. With settle on, it returns what changed on screen afterwards.

  Try: “Tap the Add button.”

  Write · press

• Clicks an item

  Clicks an item by reference or selector. The server guide lists it alongside press.

  Try: “Click the Sign in button.”

  Write · click

• Types text into a field

  Fills a text field with the text you give it. It returns what changed on screen.

  Try: “Fill the first name field with Ada.”

  Write · fill

• Presses and holds an item

  Long-presses an item by reference or selector.

  Try: “Long-press the first message in the list.”

  Write · longpress

• Moves the pointer over an item

  Hovers over an item by reference or selector.

  Try: “Hover over the Settings menu.”

  Write · hover

• Scrolls the screen

  Scrolls in a direction, or until a named item is on screen. The direction bottom runs to the end of the content.

  Try: “Scroll down until the Delete account button is visible.”

  Write · scroll

• Goes back one screen

  Navigates back in the app.

  Try: “Go back to the previous screen.”

  Write · back

• Lists simulators, emulators and connected devices

  Lists the devices the agent can control. Used when the app id or target is unknown.

  Try: “Which devices can you control?”

  Read · devices

• Lists apps installed on a device

  Lists apps on the chosen device so the agent can find the right app id instead of guessing.

  Try: “Which apps are installed on the Android emulator?”

  Read · apps

• Checks which app is in front

  Reports the state of the app on the device.

  Try: “Which app is open right now?”

  Read · appstate

• Ends the session

  Closes the session when the agent is done.

  Try: “Close the session.”

  Write · close

• Shows usage guides for the other tools

  Returns the usage guide, or the full flag reference for one tool. It is meant for unusual work, not for startup.

  Try: “Show me the help for the scroll tool.”

  Read · help

Watch out for

• The agent can tap and type in any app on the device

  On a real phone or a signed-in simulator it can send messages, change settings or buy things. Use a test device or a simulator with test accounts. github.com

• It sees everything on screen

  Snapshots and screenshots include whatever the app shows, such as messages or personal data. Close private apps before a session. github.com

• No read-only mode is documented

  The README does not describe a way to limit the server to read tools. Keep an eye on the agent while it runs. github.com

• It drives devices through local tooling and the CLI's other commands

  The README says it uses ADB on Android, XCTest on iOS and tvOS, HDC on HarmonyOS, Vega CLI on Vega devices and local helpers on macOS and Linux. The server builds its tools from the CLI command list, so the full set, which can include commands beyond the 15 listed here, is wider than this page shows. Check the help tool before giving it a device with real data. github.com

Security advisories