Mobile Simulator & Emulator MCP

A local Model Context Protocol server that lets coding agents inspect, control, and test Apple iOS Simulators and Google Android Emulators. It exposes 38 typed tools for device lifecycle, app discovery and management, UI automation, screenshots, logs, permissions, location, appearance/accessibility, and platform-specific test conditions.

It defaults to stdio and can also serve Streamable HTTP on loopback with --listen.

It uses:

• xcrun simctl for iOS Simulator lifecycle, apps, URLs, and screenshots.

• Appium WebDriverAgent for iOS UI trees and input.

• Google adb and emulator for Android lifecycle, apps, UI trees, screenshots, and input.

Physical devices are intentionally rejected.

Requirements

• macOS with Xcode and at least one installed iOS Simulator runtime.

• Android SDK Platform Tools and Android Emulator.

• Node.js 24 or newer.

• A booted simulator/emulator for interaction tools. mobile_boot_device can boot an installed target returned by mobile_list_devices.

Verify the native tools:

xcrun simctl list devices
adb devices -l
emulator -list-avds

Related MCP server: Simulator MCP

Quick start

This package is currently installed from source:

git clone https://github.com/gitsoufiane/mobile-simulator-emulator-mcp.git
cd mobile-simulator-emulator-mcp
npm ci --ignore-scripts
npm run build

Start the devices you want the agent to use:

# iOS: boot a Simulator from Xcode or the command line
open -a Simulator

# Android: list and start an installed AVD
emulator -list-avds
emulator -avd <AVD_NAME>

Then verify both native toolchains:

xcrun simctl list devices
adb devices -l
npm run live

The first iOS UI interaction builds and launches WebDriverAgent through Xcode. It can take several minutes. Later calls reuse that process until the MCP server exits or another iOS Simulator is selected.

Configure an MCP client

Any MCP client that supports local stdio servers can run this server. Build first, then add the following server definition to the client's MCP configuration, replacing the absolute path:

{
  "mcpServers": {
    "mobile": {
      "command": "node",
      "args": [
        "/absolute/path/to/mobile-simulator-emulator-mcp/dist/src/index.js"
      ]
    }
  }
}

HTTP mode (--listen)

For MCP clients that connect by URL instead of spawning a stdio process (for example flue's connectMcpServer), start the server in Streamable HTTP mode:

node dist/src/index.js --listen 3000
# ready on http://127.0.0.1:3000/mcp

The HTTP listener binds to 127.0.0.1 only and has no authentication — it is strictly a same-machine transport with the same trust model as stdio. Each MCP session gets its own server instance; sessions are cleaned up on disconnect. stdio remains the default when --listen is not given.

Restart the MCP client after changing its configuration. The first calls should be:

1. mobile_doctor with {}.

2. mobile_list_devices with {}.

3. Use the returned ios:..., android:..., or android-avd:... ID in later calls.

By default, .app and .apk installation paths are resolved beneath the MCP process working directory. Clients that launch the server from the active agent project—such as Pi's MCP adapter when no cwd is configured—therefore need no extra environment setting.

If another MCP client starts the server from a fixed directory, set that client's cwd to the mobile project. MOBILE_MCP_APP_ROOT remains available as an optional explicit override.

The server writes MCP protocol messages only to stdout. Diagnostics and WebDriverAgent logs go to stderr.

Recommended agent workflow

1. Call mobile_doctor.

2. Call mobile_list_devices and keep the returned stable ID.

3. Boot a shutdown target if needed.

4. Install and launch the app. Use mobile_list_apps to look up the exact bundle ID or package name instead of guessing, and mobile_get_foreground_app to confirm the launch landed.

5. Set permissions or simulated location needed by the scenario.

6. Call mobile_get_ui_tree before using coordinates; use mobile_take_screenshot when visual context is needed.

7. Interact with mobile_tap, mobile_swipe, mobile_type_text, and mobile_press_key.

8. Re-read the UI tree or screenshot after each state-changing action.

9. Collect mobile_get_logs when a workflow fails. Use mobile_reset_app only when a clean app state is required.

Example requests

Once connected, an agent can handle requests such as:

• “List my available simulators and emulators, then boot the shutdown iPhone.”

• “Install the Android debug APK, launch it, and take a screenshot.”

• “Open myapp://checkout/123, inspect the UI tree, and tap Continue.”

• “Set the emulator location to San Francisco and test the nearby-results screen.”

• “Collect the last five minutes of logs for com.example.app.”

• “Reset the test app, grant location permission, and rerun the onboarding flow.”

Destructive tools run without an extra MCP-side confirmation. Configure this server only for trusted agents and disposable simulator/emulator data.

Tools

Device IDs

• Booted iOS: ios:<simulator-UDID>

• Running Android: android:emulator-5554

• Shutdown Android AVD: android-avd:<AVD-name>

Always use IDs from the latest mobile_list_devices result.

Configuration

Security model

• Local stdio transport only; there is no MCP HTTP listener or remote authentication surface.

• No arbitrary shell, adb shell, or simctl spawn tool is exposed.

• Native commands use argument arrays, explicit Android serials, timeouts, and output limits.

• Physical devices, device creation/deletion, raw native-command passthrough, and arbitrary host shell access are not supported.

• Curated destructive operations—including app reset/uninstall and iOS Simulator erase—execute without per-call confirmation. The MCP client configuration is the trust boundary.

• URL schemes, device identifiers, app IDs, permissions, coordinates, environment profiles, durations, push payloads, and install paths are validated.

• WebDriverAgent is contacted only at 127.0.0.1:8100.

Treat this server like any local developer tool: it runs with the MCP client's operating-system permissions and can interact with apps displayed in test devices.

Limits

• One iOS Simulator can own the WDA process on port 8100 at a time. Android can run concurrently.

• mobile_get_ui_tree and Android input need the screen on and unlocked. A sleeping Android Emulator makes uiautomator dump fail, and the server reports that as an error rather than returning an empty tree.

• mobile_set_accessibility splits by platform: contentSize/increaseContrast are iOS-only, fontScale is Android-only.

• mobile_list_apps returns Android package names without display labels; iOS returns CFBundleDisplayName when the app declares one.

• Android input text supports printable ASCII only. iOS typing supports Unicode through WDA.

• UI trees depend on apps exposing useful accessibility labels/identifiers.

• Trees and logs over one million characters return truncated: true; treat the result as incomplete.

• Android Emulator has no official command to clear its last GPS fix; the release gate restores it to 0,0.

• Android AVD wipe is a boot-time option, so mobile_erase_device is iOS-only.

• Clipboard and cosmetic status-bar control are iOS-only in the installed official CLIs. SMS, fingerprint, network profiles, battery state, and rotation are Android-only.

• Simulated push delivery is iOS-only. Android FCM requires Firebase rather than adb.

• Apple exposes no supported simctl command for network conditioning or Face ID/Touch ID events; those remain out of scope.

• Video recording remains out of scope because it requires a long-running stream lifecycle and artifact storage policy.

• Simulator keychain manipulation, app-container host paths, iCloud sync, and device create/clone/delete/rename stay out of scope: they leak host paths or manage fleets rather than test a running device.

• MCP resources, prompts, and completions are not implemented. Device state changes too quickly to cache as resources, and every workflow here is better expressed as an explicit tool call.

• Simulators and emulators do not reproduce all real-device sensors, performance, or hardware behavior.

Troubleshooting

mobile_doctor reports a missing Android command

Set ANDROID_SDK_ROOT or ANDROID_HOME to the Android SDK directory and ensure its platform-tools and emulator directories are installed. You can also add both directories to PATH.

iOS UI calls time out on first use

WebDriverAgent may need several minutes to compile through Xcode. Set MOBILE_MCP_SHOW_XCODE_LOG=1 in the MCP environment to send detailed build logs to stderr.

App installation is rejected

The artifact must resolve beneath the agent project's working directory, or beneath MOBILE_MCP_APP_ROOT when explicitly configured. Android requires an .apk file; iOS requires a built .app directory for the Simulator architecture.

A deep link scheme is rejected

http and https are enabled by default. Add custom schemes as a comma-separated environment value, for example MOBILE_MCP_URL_SCHEMES=myapp,example-beta.

A device ID is rejected

Call mobile_list_devices again and use its exact current ID. Physical Android and Apple devices are intentionally unsupported.

Development and release checks

npm run check
npm run live
npm run live:deep
npm run live:tools
npm audit --omit=dev

npm run live:tools covers the app-discovery, appearance, accessibility, screen-info, and long-press tools, and asserts that platform and coordinate rejections still reject. It exercises Android in full and skips the iOS paths that would take over the Simulator with WebDriverAgent; set MOBILE_MCP_LIVE_WDA=1 to include them.

npm run live performs fast screenshots, UI-tree reads, and HOME input on both platforms. npm run live:deep uses the MCP stdio protocol to manipulate Android and iOS Settings, exercise a spare iOS boot/shutdown when available, set simulated locations, collect app logs, round-trip the iOS clipboard, override/clear the iOS status bar, change/restore Android battery and network profiles, rotate Android four times, remove a fingerprint touch, type and swipe, open https://example.com, verify screenshots and UI changes, exercise rejected inputs, and restore managed state.

Destructive app/device tools, permission changes, push, SMS, and install/uninstall need a disposable application/device fixture. The release gate does not run them against system apps.

Official references

The implementation and tool limits were checked against:

• MCP tool specification for strict inputs, structured results, annotations, errors, and the client-side trust model for sensitive operations.

• Apple Xcode command-line tool reference and the installed xcrun simctl help for Simulator lifecycle, logs, privacy, location, push, and screenshots.

• Android Debug Bridge, Logcat, and Emulator console for app state, permissions, logs, location, and emulator lifecycle.

See PLAN.md for the approved scope and release gates.