diff --git a/android/app/src/main/java/io/picoclaw/android/assistant/AssistantService.kt b/android/app/src/main/java/io/picoclaw/android/assistant/AssistantService.kt
index 308faeea5..fca09fe76 100644
--- a/android/app/src/main/java/io/picoclaw/android/assistant/AssistantService.kt
+++ b/android/app/src/main/java/io/picoclaw/android/assistant/AssistantService.kt
@@ -75,10 +75,12 @@ class AssistantService : LifecycleService(), SavedStateRegistryOwner {
     private val httpClient: HttpClient by inject()
     private val ttsSettingsRepo: TtsSettingsRepository by inject()
     private val screenshotSource: ScreenshotSource by inject()
+    private val deviceController: DeviceController by inject()
 
     private lateinit var serviceScope: CoroutineScope
     private lateinit var connection: AssistantConnection
     private lateinit var assistantManager: AssistantManager
+    private lateinit var toolRequestHandler: ToolRequestHandler
     private lateinit var ttsWrapper: TextToSpeechWrapper
     private lateinit var sttWrapper: SpeechRecognizerWrapper
     private lateinit var cameraCaptureManager: CameraCaptureManager
@@ -112,6 +114,20 @@ class AssistantService : LifecycleService(), SavedStateRegistryOwner {
 
         connection = AssistantConnectionImpl(httpClient)
 
+        toolRequestHandler = ToolRequestHandler(
+            context = applicationContext,
+            deviceController = deviceController,
+            onAccessibilityNeeded = { showAccessibilityGuide = true }
+        )
+        (connection as AssistantConnectionImpl).onToolRequest = { request ->
+            val response = toolRequestHandler.handle(request)
+            if (response.success) {
+                response.result ?: ""
+            } else {
+                response.error ?: "unknown error"
+            }
+        }
+
         sttWrapper = SpeechRecognizerWrapper(this)
         ttsWrapper = TextToSpeechWrapper(this, ttsSettingsRepo.ttsConfig)
         cameraCaptureManager = CameraCaptureManager(this)
diff --git a/android/app/src/main/java/io/picoclaw/android/assistant/DeviceController.kt b/android/app/src/main/java/io/picoclaw/android/assistant/DeviceController.kt
new file mode 100644
index 000000000..456527641
--- /dev/null
+++ b/android/app/src/main/java/io/picoclaw/android/assistant/DeviceController.kt
@@ -0,0 +1,89 @@
+package io.picoclaw.android.assistant
+
+import android.accessibilityservice.AccessibilityService
+import android.accessibilityservice.GestureDescription
+import android.graphics.Path
+import android.os.Bundle
+import android.view.accessibility.AccessibilityNodeInfo
+import kotlinx.coroutines.suspendCancellableCoroutine
+import kotlin.coroutines.resume
+
+class DeviceController {
+
+    @Volatile
+    private var service: AccessibilityService? = null
+
+    val isAvailable: Boolean get() = service != null
+
+    fun setService(s: AccessibilityService) {
+        service = s
+    }
+
+    fun clearService() {
+        service = null
+    }
+
+    suspend fun tap(x: Float, y: Float): Boolean {
+        val svc = service ?: return false
+        val path = Path().apply { moveTo(x, y) }
+        val stroke = GestureDescription.StrokeDescription(path, 0, 100)
+        val gesture = GestureDescription.Builder().addStroke(stroke).build()
+        return dispatchGesture(svc, gesture)
+    }
+
+    suspend fun swipe(x1: Float, y1: Float, x2: Float, y2: Float, durationMs: Long = 300): Boolean {
+        val svc = service ?: return false
+        val path = Path().apply {
+            moveTo(x1, y1)
+            lineTo(x2, y2)
+        }
+        val stroke = GestureDescription.StrokeDescription(path, 0, durationMs)
+        val gesture = GestureDescription.Builder().addStroke(stroke).build()
+        return dispatchGesture(svc, gesture)
+    }
+
+    fun pressBack(): Boolean {
+        return service?.performGlobalAction(AccessibilityService.GLOBAL_ACTION_BACK) ?: false
+    }
+
+    fun pressHome(): Boolean {
+        return service?.performGlobalAction(AccessibilityService.GLOBAL_ACTION_HOME) ?: false
+    }
+
+    fun pressRecents(): Boolean {
+        return service?.performGlobalAction(AccessibilityService.GLOBAL_ACTION_RECENTS) ?: false
+    }
+
+    fun getCurrentPackage(): String? {
+        return service?.rootInActiveWindow?.packageName?.toString()
+    }
+
+    suspend fun inputText(text: String): Boolean {
+        val svc = service ?: return false
+        val focusedNode = svc.rootInActiveWindow?.findFocus(AccessibilityNodeInfo.FOCUS_INPUT)
+            ?: return false
+        val args = Bundle().apply {
+            putCharSequence(AccessibilityNodeInfo.ACTION_ARGUMENT_SET_TEXT_CHARSEQUENCE, text)
+        }
+        return focusedNode.performAction(AccessibilityNodeInfo.ACTION_SET_TEXT, args)
+    }
+
+    private suspend fun dispatchGesture(
+        svc: AccessibilityService,
+        gesture: GestureDescription
+    ): Boolean = suspendCancellableCoroutine { cont ->
+        svc.dispatchGesture(
+            gesture,
+            object : AccessibilityService.GestureResultCallback() {
+                override fun onCompleted(gestureDescription: GestureDescription?) {
+                    cont.resume(true)
+                }
+
+                override fun onCancelled(gestureDescription: GestureDescription?) {
+                    cont.resume(false)
+                }
+            },
+            null
+        )
+    }
+}
diff --git a/android/app/src/main/java/io/picoclaw/android/assistant/PicoClawAccessibilityService.kt b/android/app/src/main/java/io/picoclaw/android/assistant/PicoClawAccessibilityService.kt
index b49c7473d..1e8149d5b 100644
--- a/android/app/src/main/java/io/picoclaw/android/assistant/PicoClawAccessibilityService.kt
+++ b/android/app/src/main/java/io/picoclaw/android/assistant/PicoClawAccessibilityService.kt
@@ -7,19 +7,22 @@ import org.koin.android.ext.android.inject
 class PicoClawAccessibilityService : AccessibilityService() {
 
     private val screenshotSource: AccessibilityScreenshotSource by inject()
+    private val deviceController: DeviceController by inject()
 
     override fun onServiceConnected() {
         super.onServiceConnected()
         screenshotSource.setService(this)
+        deviceController.setService(this)
     }
 
     override fun onDestroy() {
         screenshotSource.clearService()
+        deviceController.clearService()
         super.onDestroy()
     }
 
     override fun onAccessibilityEvent(event: AccessibilityEvent?) {
-        // No-op: used only for screenshot capture
+        // No-op
     }
 
     override fun onInterrupt() {
diff --git a/android/app/src/main/java/io/picoclaw/android/assistant/ToolRequestHandler.kt b/android/app/src/main/java/io/picoclaw/android/assistant/ToolRequestHandler.kt
new file mode 100644
index 000000000..2bfc7e432
--- /dev/null
+++ b/android/app/src/main/java/io/picoclaw/android/assistant/ToolRequestHandler.kt
@@ -0,0 +1,260 @@
+package io.picoclaw.android.assistant
+
+import android.content.Context
+import android.content.Intent
+import android.content.pm.ApplicationInfo
+import android.content.pm.PackageManager
+import android.net.Uri
+import android.os.Build
+import android.util.Log
+import io.picoclaw.android.core.data.remote.dto.ToolRequest
+import io.picoclaw.android.core.data.remote.dto.ToolResponse
+import kotlinx.serialization.json.JsonObject
+import kotlinx.serialization.json.JsonPrimitive
+import kotlinx.serialization.json.contentOrNull
+import kotlinx.serialization.json.doubleOrNull
+import kotlinx.serialization.json.jsonPrimitive
+import kotlinx.serialization.json.longOrNull
+
+class ToolRequestHandler(
+    private val context: Context,
+    private val deviceController: DeviceController,
+    private val onAccessibilityNeeded: () -> Unit
+) {
+
+    suspend fun handle(request: ToolRequest): ToolResponse {
+        return try {
+            when (request.action) {
+                "list_apps" -> handleListApps(request)
+                "app_info" -> handleAppInfo(request)
+                "launch_app" -> handleLaunchApp(request)
+                "current_activity" -> handleCurrentActivity(request)
+                "tap" -> handleTap(request)
+                "swipe" -> handleSwipe(request)
+                "text" -> handleText(request)
+                "keyevent" -> handleKeyEvent(request)
+                "broadcast" -> handleBroadcast(request)
+                "intent" -> handleIntent(request)
+                else -> ToolResponse(
+                    requestId = request.requestId,
+                    success = false,
+                    error = "Unknown action: ${request.action}"
+                )
+            }
+        } catch (e: Exception) {
+            Log.e(TAG, "Error handling tool request: ${request.action}", e)
+            ToolResponse(
+                requestId = request.requestId,
+                success = false,
+                error = "Error: ${e.message}"
+            )
+        }
+    }
+
+    private fun requireAccessibility(request: ToolRequest): ToolResponse? {
+        if (!deviceController.isAvailable) {
+            onAccessibilityNeeded()
+            return ToolResponse(
+                requestId = request.requestId,
+                success = false,
+                error = "accessibility_required"
+            )
+        }
+        return null
+    }
+
+    private fun handleListApps(request: ToolRequest): ToolResponse {
+        val pm = context.packageManager
+        val apps = pm.getInstalledApplications(PackageManager.GET_META_DATA)
+            .filter { pm.getLaunchIntentForPackage(it.packageName) != null }
+            .map { app ->
+                val label = pm.getApplicationLabel(app).toString()
+                "${label} (${app.packageName})"
+            }
+            .sorted()
+
+        return ToolResponse(
+            requestId = request.requestId,
+            success = true,
+            result = "Installed apps (${apps.size}):\n${apps.joinToString("\n")}"
+        )
+    }
+
+    private fun handleAppInfo(request: ToolRequest): ToolResponse {
+        val packageName = request.params?.get("package_name")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "package_name required")
+
+        val pm = context.packageManager
+        return try {
+            val info = if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU) {
+                pm.getPackageInfo(packageName, PackageManager.PackageInfoFlags.of(0))
+            } else {
+                @Suppress("DEPRECATION")
+                pm.getPackageInfo(packageName, 0)
+            }
+            val appInfo = info.applicationInfo
+            val label = appInfo?.let { pm.getApplicationLabel(it).toString() } ?: packageName
+            val isSystem = appInfo?.flags?.and(ApplicationInfo.FLAG_SYSTEM) != 0
+
+            val sb = StringBuilder()
+            sb.appendLine("App: $label")
+            sb.appendLine("Package: $packageName")
+            sb.appendLine("Version: ${info.versionName ?: "unknown"}")
+            sb.appendLine("System app: $isSystem")
+            if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.P) {
+                sb.appendLine("Version code: ${info.longVersionCode}")
+            }
+
+            ToolResponse(request.requestId, true, result = sb.toString())
+        } catch (e: PackageManager.NameNotFoundException) {
+            ToolResponse(request.requestId, false, error = "Package not found: $packageName")
+        }
+    }
+
+    private fun handleLaunchApp(request: ToolRequest): ToolResponse {
+        val packageName = request.params?.get("package_name")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "package_name required")
+
+        val intent = context.packageManager.getLaunchIntentForPackage(packageName)
+            ?: return ToolResponse(request.requestId, false, error = "No launch intent for $packageName")
+
+        intent.addFlags(Intent.FLAG_ACTIVITY_NEW_TASK)
+        context.startActivity(intent)
+        return ToolResponse(request.requestId, true, result = "Launched $packageName")
+    }
+
+    private suspend fun handleCurrentActivity(request: ToolRequest): ToolResponse {
+        requireAccessibility(request)?.let { return it }
+
+        val pkg = deviceController.getCurrentPackage()
+            ?: return ToolResponse(request.requestId, false, error = "Could not get current activity")
+
+        return ToolResponse(request.requestId, true, result = "Current package: $pkg")
+    }
+
+    private suspend fun handleTap(request: ToolRequest): ToolResponse {
+        requireAccessibility(request)?.let { return it }
+
+        val x = request.params?.get("x")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "x coordinate required")
+        val y = request.params?.get("y")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "y coordinate required")
+
+        val success = deviceController.tap(x, y)
+        return ToolResponse(
+            request.requestId, success,
+            result = if (success) "Tapped at ($x, $y)" else null,
+            error = if (!success) "Tap failed" else null
+        )
+    }
+
+    private suspend fun handleSwipe(request: ToolRequest): ToolResponse {
+        requireAccessibility(request)?.let { return it }
+
+        val x = request.params?.get("x")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "x coordinate required")
+        val y = request.params?.get("y")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "y coordinate required")
+        val x2 = request.params?.get("x2")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "x2 coordinate required")
+        val y2 = request.params?.get("y2")?.jsonPrimitive?.doubleOrNull?.toFloat()
+            ?: return ToolResponse(request.requestId, false, error = "y2 coordinate required")
+        val durationMs = request.params?.get("duration_ms")?.jsonPrimitive?.longOrNull ?: 300L
+
+        val success = deviceController.swipe(x, y, x2, y2, durationMs)
+        return ToolResponse(
+            request.requestId, success,
+            result = if (success) "Swiped from ($x,$y) to ($x2,$y2)" else null,
+            error = if (!success) "Swipe failed" else null
+        )
+    }
+
+    private suspend fun handleText(request: ToolRequest): ToolResponse {
+        requireAccessibility(request)?.let { return it }
+
+        val text = request.params?.get("text")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "text required")
+
+        val success = deviceController.inputText(text)
+        return ToolResponse(
+            request.requestId, success,
+            result = if (success) "Text input: $text" else null,
+            error = if (!success) "Text input failed (no focused input field?)" else null
+        )
+    }
+
+    private fun handleKeyEvent(request: ToolRequest): ToolResponse {
+        if (!deviceController.isAvailable) {
+            onAccessibilityNeeded()
+            return ToolResponse(request.requestId, false, error = "accessibility_required")
+        }
+
+        val key = request.params?.get("key")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "key required")
+
+        val success = when (key) {
+            "back" -> deviceController.pressBack()
+            "home" -> deviceController.pressHome()
+            "recents" -> deviceController.pressRecents()
+            else -> return ToolResponse(request.requestId, false, error = "Unknown key: $key")
+        }
+        return ToolResponse(
+            request.requestId, success,
+            result = if (success) "Key pressed: $key" else null,
+            error = if (!success) "Key event failed" else null
+        )
+    }
+
+    private fun handleBroadcast(request: ToolRequest): ToolResponse {
+        val action = request.params?.get("intent_action")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "intent_action required")
+
+        val intent = Intent(action)
+        applyExtras(intent, request.params)
+        context.sendBroadcast(intent)
+        return ToolResponse(request.requestId, true, result = "Broadcast sent: $action")
+    }
+
+    private fun handleIntent(request: ToolRequest): ToolResponse {
+        val action = request.params?.get("intent_action")?.jsonPrimitive?.contentOrNull
+            ?: return ToolResponse(request.requestId, false, error = "intent_action required")
+
+        val intent = Intent(action).addFlags(Intent.FLAG_ACTIVITY_NEW_TASK)
+
+        request.params?.get("intent_data")?.jsonPrimitive?.contentOrNull?.let {
+            intent.data = Uri.parse(it)
+        }
+        request.params?.get("intent_package")?.jsonPrimitive?.contentOrNull?.let {
+            intent.setPackage(it)
+        }
+        request.params?.get("intent_type")?.jsonPrimitive?.contentOrNull?.let {
+            intent.type = it
+        }
+        applyExtras(intent, request.params)
+
+        return try {
+            context.startActivity(intent)
+            ToolResponse(request.requestId, true, result = "Intent started: $action")
+        } catch (e: Exception) {
+            ToolResponse(request.requestId, false, error = "Failed to start intent: ${e.message}")
+        }
+    }
+
+    private fun applyExtras(intent: Intent, params: JsonObject?) {
+        val extras = params?.get("intent_extras") as? JsonObject ?: return
+        for ((key, value) in extras) {
+            val prim = value as? JsonPrimitive ?: continue
+            when {
+                prim.isString -> intent.putExtra(key, prim.content)
+                prim.content.toBooleanStrictOrNull() != null ->
+                    intent.putExtra(key, prim.content.toBooleanStrict())
+                prim.longOrNull != null -> intent.putExtra(key, prim.longOrNull!!)
+                prim.doubleOrNull != null -> intent.putExtra(key, prim.doubleOrNull!!)
+            }
+        }
+    }
+
+    companion object {
+        private const val TAG = "ToolRequestHandler"
+    }
+}
diff --git a/android/app/src/main/java/io/picoclaw/android/di/AppModule.kt b/android/app/src/main/java/io/picoclaw/android/di/AppModule.kt
index a0bec7d64..35bebb022 100644
--- a/android/app/src/main/java/io/picoclaw/android/di/AppModule.kt
+++ b/android/app/src/main/java/io/picoclaw/android/di/AppModule.kt
@@ -6,6 +6,7 @@ import io.ktor.client.engine.okhttp.OkHttp
 import io.ktor.client.plugins.websocket.WebSockets
 import io.picoclaw.android.core.data.local.AppDatabase
 import io.picoclaw.android.assistant.AccessibilityScreenshotSource
+import io.picoclaw.android.assistant.DeviceController
 import io.picoclaw.android.core.data.local.ImageFileStorage
 import io.picoclaw.android.feature.chat.voice.ScreenshotSource
 import io.picoclaw.android.core.data.remote.WebSocketClient
@@ -96,6 +97,9 @@ val appModule = module {
     single { AccessibilityScreenshotSource() }
     single<ScreenshotSource> { get<AccessibilityScreenshotSource>() }
 
+    // Device Controller (for Android tool)
+    single { DeviceController() }
+
     // Voice
     factory { SpeechRecognizerWrapper(androidContext()) }
     single { TextToSpeechWrapper(androidContext(), get<TtsSettingsRepository>().ttsConfig) }
diff --git a/android/app/src/main/res/xml/accessibility_service_config.xml b/android/app/src/main/res/xml/accessibility_service_config.xml
index 429f45e65..9561a6a3a 100644
--- a/android/app/src/main/res/xml/accessibility_service_config.xml
+++ b/android/app/src/main/res/xml/accessibility_service_config.xml
@@ -1,5 +1,7 @@
 <?xml version="1.0" encoding="utf-8"?>
 <accessibility-service xmlns:android="http://schemas.android.com/apk/res/android"
     android:canTakeScreenshot="true"
-    android:canRetrieveWindowContent="false"
+    android:canRetrieveWindowContent="true"
+    android:canPerformGestures="true"
+    android:accessibilityFlags="flagReportViewIds"
     android:description="@string/accessibility_service_description" />
diff --git a/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolRequest.kt b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolRequest.kt
new file mode 100644
index 000000000..779bc1887
--- /dev/null
+++ b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolRequest.kt
@@ -0,0 +1,12 @@
+package io.picoclaw.android.core.data.remote.dto
+
+import kotlinx.serialization.SerialName
+import kotlinx.serialization.Serializable
+import kotlinx.serialization.json.JsonObject
+
+@Serializable
+data class ToolRequest(
+    @SerialName("request_id") val requestId: String,
+    val action: String,
+    val params: JsonObject? = null
+)
diff --git a/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolResponse.kt b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolResponse.kt
new file mode 100644
index 000000000..e335a145d
--- /dev/null
+++ b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/ToolResponse.kt
@@ -0,0 +1,12 @@
+package io.picoclaw.android.core.data.remote.dto
+
+import kotlinx.serialization.SerialName
+import kotlinx.serialization.Serializable
+
+@Serializable
+data class ToolResponse(
+    @SerialName("request_id") val requestId: String,
+    val success: Boolean,
+    val result: String? = null,
+    val error: String? = null
+)
diff --git a/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/WsIncoming.kt b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/WsIncoming.kt
index ebc2add90..4995da5e3 100644
--- a/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/WsIncoming.kt
+++ b/android/core/data/src/main/java/io/picoclaw/android/core/data/remote/dto/WsIncoming.kt
@@ -8,5 +8,7 @@ data class WsIncoming(
     val content: String,
     @SerialName("sender_id") val senderId: String? = null,
     val images: List<String>? = null,
-    @SerialName("input_mode") val inputMode: String? = null
+    @SerialName("input_mode") val inputMode: String? = null,
+    val type: String? = null,
+    @SerialName("request_id") val requestId: String? = null
 )
diff --git a/android/core/data/src/main/java/io/picoclaw/android/core/data/repository/AssistantConnectionImpl.kt b/android/core/data/src/main/java/io/picoclaw/android/core/data/repository/AssistantConnectionImpl.kt
index 289a0566b..3640da070 100644
--- a/android/core/data/src/main/java/io/picoclaw/android/core/data/repository/AssistantConnectionImpl.kt
+++ b/android/core/data/src/main/java/io/picoclaw/android/core/data/repository/AssistantConnectionImpl.kt
@@ -1,7 +1,9 @@
 package io.picoclaw.android.core.data.repository
 
+import android.util.Log
 import io.ktor.client.HttpClient
 import io.picoclaw.android.core.data.remote.WebSocketClient
+import io.picoclaw.android.core.data.remote.dto.ToolRequest
 import io.picoclaw.android.core.data.remote.dto.WsIncoming
 import io.picoclaw.android.core.domain.model.AssistantMessage
 import io.picoclaw.android.core.domain.model.ConnectionState
@@ -17,8 +19,11 @@ import kotlinx.coroutines.flow.StateFlow
 import kotlinx.coroutines.flow.asSharedFlow
 import kotlinx.coroutines.flow.asStateFlow
 import kotlinx.coroutines.launch
+import kotlinx.serialization.json.Json
 import java.util.UUID
 
+typealias ToolRequestCallback = suspend (ToolRequest) -> String
+
 class AssistantConnectionImpl(
     private val httpClient: HttpClient
 ) : AssistantConnection {
@@ -26,6 +31,7 @@ class AssistantConnectionImpl(
     private val scope = CoroutineScope(SupervisorJob() + Dispatchers.IO)
     private val clientId = UUID.randomUUID().toString()
     private val wsClient = WebSocketClient(httpClient, scope, clientId, "assistant")
+    private val json = Json { ignoreUnknownKeys = true }
 
     private val _messages = MutableSharedFlow<AssistantMessage>(extraBufferCapacity = 64)
     override val messages: SharedFlow<AssistantMessage> = _messages.asSharedFlow()
@@ -35,12 +41,15 @@ class AssistantConnectionImpl(
 
     override val connectionState: StateFlow<ConnectionState> = wsClient.connectionState
 
+    var onToolRequest: ToolRequestCallback? = null
+
     init {
         scope.launch {
             wsClient.incomingMessages.collect { dto ->
                 when (dto.type) {
                     "status" -> _statusText.value = dto.content
                     "status_end" -> _statusText.value = null
+                    "tool_request" -> handleToolRequest(dto.content)
                     else -> {
                         _statusText.value = null
                         _messages.emit(AssistantMessage(content = dto.content, type = dto.type))
@@ -50,6 +59,29 @@ class AssistantConnectionImpl(
         }
     }
 
+    private fun handleToolRequest(content: String) {
+        scope.launch {
+            try {
+                val request = json.decodeFromString<ToolRequest>(content)
+                val callback = onToolRequest
+                val resultContent = if (callback != null) {
+                    callback(request)
+                } else {
+                    "error: tool request handler not configured"
+                }
+
+                val response = WsIncoming(
+                    content = resultContent,
+                    type = "tool_response",
+                    requestId = request.requestId
+                )
+                wsClient.send(response)
+            } catch (e: Exception) {
+                Log.e(TAG, "Failed to handle tool request", e)
+            }
+        }
+    }
+
     override fun connect(wsUrl: String) {
         wsClient.wsUrl = wsUrl
         wsClient.connect()
@@ -68,4 +100,8 @@ class AssistantConnectionImpl(
         )
         wsClient.send(dto)
     }
+
+    companion object {
+        private const val TAG = "AssistantConnectionImpl"
+    }
 }
diff --git a/pkg/agent/loop.go b/pkg/agent/loop.go
index 1dba75467..36bb381ff 100644
--- a/pkg/agent/loop.go
+++ b/pkg/agent/loop.go
@@ -107,6 +107,21 @@ func createToolRegistry(workspace string, restrict bool, cfg *config.Config, msg
 		registry.Register(tools.NewSPITool())
 	}
 
+	// Android device control tool (assistant mode only)
+	if cfg.Tools.Android.Enabled {
+		androidTool := tools.NewAndroidTool()
+		androidTool.SetSendCallback(func(channel, chatID, content, msgType string) error {
+			msgBus.PublishOutbound(bus.OutboundMessage{
+				Channel: channel,
+				ChatID:  chatID,
+				Content: content,
+				Type:    msgType,
+			})
+			return nil
+		})
+		registry.Register(androidTool)
+	}
+
 	// Message tool - available to both agent and subagent
 	// Subagent uses it to communicate directly with user
 	messageTool := tools.NewMessageTool()
@@ -510,7 +525,7 @@ func (al *AgentLoop) runAgentLoop(ctx context.Context, opts processOptions) (str
 	}
 
 	// 1. Update tool contexts
-	al.updateToolContexts(opts.Channel, opts.ChatID)
+	al.updateToolContexts(opts.Channel, opts.ChatID, opts.Metadata)
 
 	// 2. Build messages (skip history for heartbeat)
 	var history []providers.Message
@@ -959,7 +974,7 @@ func (al *AgentLoop) runLLMIteration(ctx context.Context, messages []providers.M
 }
 
 // updateToolContexts updates the context for tools that need channel/chatID info.
-func (al *AgentLoop) updateToolContexts(channel, chatID string) {
+func (al *AgentLoop) updateToolContexts(channel, chatID string, metadata map[string]string) {
 	// Use ContextualTool interface instead of type assertions
 	if tool, ok := al.tools.Get("message"); ok {
 		if mt, ok := tool.(tools.ContextualTool); ok {
@@ -976,6 +991,14 @@ func (al *AgentLoop) updateToolContexts(channel, chatID string) {
 			st.SetContext(channel, chatID)
 		}
 	}
+	if tool, ok := al.tools.Get("android"); ok {
+		if at, ok := tool.(*tools.AndroidTool); ok {
+			at.SetContext(channel, chatID)
+			if metadata != nil {
+				at.SetClientType(metadata["client_type"])
+			}
+		}
+	}
 }
 
 // maybeSummarize triggers summarization if the session history exceeds thresholds.
diff --git a/pkg/agent/status.go b/pkg/agent/status.go
index 89e7ff0f7..86a628093 100644
--- a/pkg/agent/status.go
+++ b/pkg/agent/status.go
@@ -56,6 +56,8 @@ func statusLabel(toolName string, args map[string]interface{}) string {
 			return fmt.Sprintf("サブタスク実行中...（%s）", truncLabel(l, 20))
 		}
 		return "サブタスク実行中..."
+	case "android":
+		return androidStatusLabel(args)
 	case "mcp":
 		return mcpStatusLabel(args)
 	case "i2c":
@@ -116,6 +118,42 @@ func cronStatusLabel(args map[string]interface{}) string {
 	}
 }
 
+func androidStatusLabel(args map[string]interface{}) string {
+	switch strArg(args, "action") {
+	case "list_apps":
+		return "アプリ一覧取得中..."
+	case "app_info":
+		if p := strArg(args, "package_name"); p != "" {
+			return fmt.Sprintf("アプリ情報取得中...（%s）", truncLabel(p, 25))
+		}
+		return "アプリ情報取得中..."
+	case "launch_app":
+		if p := strArg(args, "package_name"); p != "" {
+			return fmt.Sprintf("アプリ起動中...（%s）", truncLabel(p, 25))
+		}
+		return "アプリ起動中..."
+	case "current_activity":
+		return "画面情報取得中..."
+	case "tap":
+		return "タップ中..."
+	case "swipe":
+		return "スワイプ中..."
+	case "text":
+		return "テキスト入力中..."
+	case "keyevent":
+		if k := strArg(args, "key"); k != "" {
+			return fmt.Sprintf("キー操作中...（%s）", k)
+		}
+		return "キー操作中..."
+	case "broadcast":
+		return "ブロードキャスト送信中..."
+	case "intent":
+		return "インテント送信中..."
+	default:
+		return "デバイス操作中..."
+	}
+}
+
 func mcpStatusLabel(args map[string]interface{}) string {
 	switch strArg(args, "action") {
 	case "mcp_list":
diff --git a/pkg/channels/websocket.go b/pkg/channels/websocket.go
index a9b988a54..0dc4bad6b 100644
--- a/pkg/channels/websocket.go
+++ b/pkg/channels/websocket.go
@@ -13,6 +13,7 @@ import (
 	"github.com/sipeed/picoclaw/pkg/bus"
 	"github.com/sipeed/picoclaw/pkg/config"
 	"github.com/sipeed/picoclaw/pkg/logger"
+	"github.com/sipeed/picoclaw/pkg/tools"
 )
 
 // wsIncoming is the JSON message sent from APK to picoclaw.
@@ -21,6 +22,8 @@ type wsIncoming struct {
 	SenderID  string   `json:"sender_id,omitempty"`
 	Images    []string `json:"images,omitempty"`
 	InputMode string   `json:"input_mode,omitempty"`
+	Type      string   `json:"type,omitempty"`       // "tool_response" for device tool responses
+	RequestID string   `json:"request_id,omitempty"` // correlates with tool_request
 }
 
 // wsOutgoing is the JSON message sent from picoclaw to APK.
@@ -283,6 +286,12 @@ func (c *WebSocketChannel) readPump(conn *websocket.Conn, clientID, chatID, clie
 			continue
 		}
 
+		// Intercept tool_response messages: deliver to ResponseWaiter, skip inbound bus.
+		if incoming.Type == "tool_response" && incoming.RequestID != "" {
+			tools.DeviceResponseWaiter.Deliver(incoming.RequestID, incoming.Content)
+			continue
+		}
+
 		// Use sender_id from message if provided, otherwise use clientID.
 		senderID := clientID
 		if incoming.SenderID != "" {
diff --git a/pkg/config/config.go b/pkg/config/config.go
index bab5508a9..ee84b0f0c 100644
--- a/pkg/config/config.go
+++ b/pkg/config/config.go
@@ -253,12 +253,17 @@ type MCPServerConfig struct {
 	IdleTimeout int    `json:"idle_timeout,omitempty"` // seconds, default 300
 }
 
+type AndroidToolsConfig struct {
+	Enabled bool `json:"enabled" env:"PICOCLAW_TOOLS_ANDROID_ENABLED"`
+}
+
 type ToolsConfig struct {
-	Web  WebToolsConfig             `json:"web"`
-	Exec ExecToolsConfig            `json:"exec"`
-	I2C  I2CToolsConfig             `json:"i2c"`
-	SPI  SPIToolsConfig             `json:"spi"`
-	MCP  map[string]MCPServerConfig `json:"mcp,omitempty"`
+	Web     WebToolsConfig             `json:"web"`
+	Exec    ExecToolsConfig            `json:"exec"`
+	I2C     I2CToolsConfig             `json:"i2c"`
+	SPI     SPIToolsConfig             `json:"spi"`
+	Android AndroidToolsConfig         `json:"android"`
+	MCP     map[string]MCPServerConfig `json:"mcp,omitempty"`
 }
 
 func DefaultConfig() *Config {
@@ -374,6 +379,9 @@ func DefaultConfig() *Config {
 			SPI: SPIToolsConfig{
 				Enabled: false,
 			},
+			Android: AndroidToolsConfig{
+				Enabled: false,
+			},
 			Web: WebToolsConfig{
 				Brave: BraveConfig{
 					Enabled:    false,
diff --git a/pkg/tools/android.go b/pkg/tools/android.go
new file mode 100644
index 000000000..dc5bb3964
--- /dev/null
+++ b/pkg/tools/android.go
@@ -0,0 +1,334 @@
+package tools
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"regexp"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+)
+
+const androidToolTimeout = 15 * time.Second
+
+var (
+	packageNameRe  = regexp.MustCompile(`^[a-zA-Z][a-zA-Z0-9_]*(\.[a-zA-Z][a-zA-Z0-9_]*)*$`)
+	intentActionRe = regexp.MustCompile(`^[a-zA-Z][a-zA-Z0-9_.]*$`)
+)
+
+// SendCallbackWithType is like SendCallback but includes a message type field.
+type SendCallbackWithType func(channel, chatID, content, msgType string) error
+
+// toolRequest is the JSON payload sent to the Android device via WebSocket.
+type toolRequest struct {
+	RequestID string                 `json:"request_id"`
+	Action    string                 `json:"action"`
+	Params    map[string]interface{} `json:"params,omitempty"`
+}
+
+type AndroidTool struct {
+	sendCallback SendCallbackWithType
+	channel      string
+	chatID       string
+	clientType   string
+}
+
+func NewAndroidTool() *AndroidTool {
+	return &AndroidTool{}
+}
+
+func (t *AndroidTool) Name() string { return "android" }
+
+func (t *AndroidTool) Description() string {
+	return `Control the Android device. Available actions:
+- list_apps: List installed apps
+- app_info: Get app details (requires package_name)
+- launch_app: Launch an app (requires package_name)
+- current_activity: Get the currently active app/window
+- tap: Tap a screen coordinate (requires x, y)
+- swipe: Swipe between coordinates (requires x, y, x2, y2; optional duration_ms)
+- text: Input text into the focused field (requires text)
+- keyevent: Press a key (requires key: back/home/recents)
+- broadcast: Send a broadcast intent (requires intent_action; optional intent_extras)
+- intent: Start an activity via intent (requires intent_action; optional intent_data, intent_package, intent_type, intent_extras)
+
+Only available in assistant mode.`
+}
+
+func (t *AndroidTool) Parameters() map[string]interface{} {
+	return map[string]interface{}{
+		"type": "object",
+		"properties": map[string]interface{}{
+			"action": map[string]interface{}{
+				"type": "string",
+				"enum": []string{
+					"list_apps", "app_info", "launch_app", "current_activity",
+					"tap", "swipe", "text", "keyevent",
+					"broadcast", "intent",
+				},
+				"description": "The device action to perform",
+			},
+			"package_name": map[string]interface{}{
+				"type":        "string",
+				"description": "Android package name (for app_info, launch_app)",
+			},
+			"x": map[string]interface{}{
+				"type":        "number",
+				"description": "X coordinate (for tap, swipe start)",
+			},
+			"y": map[string]interface{}{
+				"type":        "number",
+				"description": "Y coordinate (for tap, swipe start)",
+			},
+			"x2": map[string]interface{}{
+				"type":        "number",
+				"description": "End X coordinate (for swipe)",
+			},
+			"y2": map[string]interface{}{
+				"type":        "number",
+				"description": "End Y coordinate (for swipe)",
+			},
+			"duration_ms": map[string]interface{}{
+				"type":        "integer",
+				"description": "Swipe duration in milliseconds (default 300)",
+			},
+			"text": map[string]interface{}{
+				"type":        "string",
+				"description": "Text to input (for text action)",
+			},
+			"key": map[string]interface{}{
+				"type":        "string",
+				"enum":        []string{"back", "home", "recents"},
+				"description": "Key to press (for keyevent action)",
+			},
+			"intent_action": map[string]interface{}{
+				"type":        "string",
+				"description": "Intent action string (for broadcast, intent)",
+			},
+			"intent_data": map[string]interface{}{
+				"type":        "string",
+				"description": "Intent data URI (for intent)",
+			},
+			"intent_package": map[string]interface{}{
+				"type":        "string",
+				"description": "Target package for intent (for intent)",
+			},
+			"intent_type": map[string]interface{}{
+				"type":        "string",
+				"description": "MIME type for intent (for intent)",
+			},
+			"intent_extras": map[string]interface{}{
+				"type":        "object",
+				"description": "Extra key-value pairs for broadcast/intent",
+			},
+		},
+		"required": []string{"action"},
+	}
+}
+
+func (t *AndroidTool) SetContext(channel, chatID string) {
+	t.channel = channel
+	t.chatID = chatID
+}
+
+func (t *AndroidTool) SetClientType(clientType string) {
+	t.clientType = clientType
+}
+
+func (t *AndroidTool) SetSendCallback(cb SendCallbackWithType) {
+	t.sendCallback = cb
+}
+
+func (t *AndroidTool) Execute(ctx context.Context, args map[string]interface{}) *ToolResult {
+	if t.clientType == "" {
+		return ErrorResult("android tool: not available (no assistant session)")
+	}
+	if t.clientType != "assistant" {
+		return ErrorResult("android tool is only available in assistant mode")
+	}
+	if t.sendCallback == nil {
+		return ErrorResult("android tool: send callback not configured")
+	}
+	if t.channel == "" || t.chatID == "" {
+		return ErrorResult("android tool: no active channel context")
+	}
+
+	action, _ := args["action"].(string)
+	if action == "" {
+		return ErrorResult("action is required")
+	}
+
+	params, err := t.validateAndBuildParams(action, args)
+	if err != nil {
+		return ErrorResult(err.Error())
+	}
+
+	return t.sendAndWait(ctx, action, params)
+}
+
+func (t *AndroidTool) validateAndBuildParams(action string, args map[string]interface{}) (map[string]interface{}, error) {
+	params := make(map[string]interface{})
+
+	switch action {
+	case "list_apps":
+		// No params needed
+
+	case "app_info", "launch_app":
+		pkg, _ := args["package_name"].(string)
+		if pkg == "" {
+			return nil, fmt.Errorf("%s requires package_name", action)
+		}
+		if !packageNameRe.MatchString(pkg) {
+			return nil, fmt.Errorf("invalid package_name: %s", pkg)
+		}
+		params["package_name"] = pkg
+
+	case "current_activity":
+		// No params needed
+
+	case "tap":
+		x, xOk := toFloat64(args["x"])
+		y, yOk := toFloat64(args["y"])
+		if !xOk || !yOk {
+			return nil, fmt.Errorf("tap requires x and y coordinates")
+		}
+		params["x"] = x
+		params["y"] = y
+
+	case "swipe":
+		x, xOk := toFloat64(args["x"])
+		y, yOk := toFloat64(args["y"])
+		x2, x2Ok := toFloat64(args["x2"])
+		y2, y2Ok := toFloat64(args["y2"])
+		if !xOk || !yOk || !x2Ok || !y2Ok {
+			return nil, fmt.Errorf("swipe requires x, y, x2, y2 coordinates")
+		}
+		params["x"] = x
+		params["y"] = y
+		params["x2"] = x2
+		params["y2"] = y2
+		if dur, ok := toFloat64(args["duration_ms"]); ok {
+			params["duration_ms"] = int(dur)
+		}
+
+	case "text":
+		text, _ := args["text"].(string)
+		if text == "" {
+			return nil, fmt.Errorf("text action requires text parameter")
+		}
+		params["text"] = text
+
+	case "keyevent":
+		key, _ := args["key"].(string)
+		if key == "" {
+			return nil, fmt.Errorf("keyevent requires key parameter")
+		}
+		switch key {
+		case "back", "home", "recents":
+			// valid
+		default:
+			return nil, fmt.Errorf("invalid key: %s (must be back, home, or recents)", key)
+		}
+		params["key"] = key
+
+	case "broadcast":
+		intentAction, _ := args["intent_action"].(string)
+		if intentAction == "" {
+			return nil, fmt.Errorf("broadcast requires intent_action")
+		}
+		if !intentActionRe.MatchString(intentAction) {
+			return nil, fmt.Errorf("invalid intent_action: %s", intentAction)
+		}
+		params["intent_action"] = intentAction
+		if extras, ok := args["intent_extras"].(map[string]interface{}); ok {
+			params["intent_extras"] = extras
+		}
+
+	case "intent":
+		intentAction, _ := args["intent_action"].(string)
+		if intentAction == "" {
+			return nil, fmt.Errorf("intent requires intent_action")
+		}
+		if !intentActionRe.MatchString(intentAction) {
+			return nil, fmt.Errorf("invalid intent_action: %s", intentAction)
+		}
+		params["intent_action"] = intentAction
+		if data, ok := args["intent_data"].(string); ok && data != "" {
+			params["intent_data"] = data
+		}
+		if pkg, ok := args["intent_package"].(string); ok && pkg != "" {
+			if !packageNameRe.MatchString(pkg) {
+				return nil, fmt.Errorf("invalid intent_package: %s", pkg)
+			}
+			params["intent_package"] = pkg
+		}
+		if mimeType, ok := args["intent_type"].(string); ok && mimeType != "" {
+			params["intent_type"] = mimeType
+		}
+		if extras, ok := args["intent_extras"].(map[string]interface{}); ok {
+			params["intent_extras"] = extras
+		}
+
+	default:
+		return nil, fmt.Errorf("unknown action: %s", action)
+	}
+
+	return params, nil
+}
+
+func (t *AndroidTool) sendAndWait(ctx context.Context, action string, params map[string]interface{}) *ToolResult {
+	requestID := uuid.New().String()
+
+	req := toolRequest{
+		RequestID: requestID,
+		Action:    action,
+		Params:    params,
+	}
+
+	reqJSON, err := json.Marshal(req)
+	if err != nil {
+		return ErrorResult(fmt.Sprintf("failed to marshal tool request: %v", err))
+	}
+
+	// Register waiter before sending to avoid race
+	respCh := DeviceResponseWaiter.Register(requestID)
+
+	if err := t.sendCallback(t.channel, t.chatID, string(reqJSON), "tool_request"); err != nil {
+		DeviceResponseWaiter.Cleanup(requestID)
+		return ErrorResult(fmt.Sprintf("failed to send tool request: %v", err))
+	}
+
+	// Wait for response with timeout
+	select {
+	case content := <-respCh:
+		// Check if the response indicates accessibility_required
+		if strings.HasPrefix(content, "accessibility_required") {
+			return &ToolResult{
+				ForUser: "この機能にはユーザー補助の設定が必要です",
+				ForLLM:  "accessibility_required: The accessibility service is not enabled. The settings dialog has been shown to the user. Do not retry automatically - wait for the user to enable the service and try again.",
+			}
+		}
+		return SilentResult(content)
+	case <-time.After(androidToolTimeout):
+		DeviceResponseWaiter.Cleanup(requestID)
+		return ErrorResult("android tool request timed out (15s)")
+	case <-ctx.Done():
+		DeviceResponseWaiter.Cleanup(requestID)
+		return ErrorResult("android tool request cancelled")
+	}
+}
+
+// toFloat64 extracts a float64 from an interface{} (handles both float64 and int from JSON).
+func toFloat64(v interface{}) (float64, bool) {
+	switch n := v.(type) {
+	case float64:
+		return n, true
+	case int:
+		return float64(n), true
+	case int64:
+		return float64(n), true
+	}
+	return 0, false
+}
diff --git a/pkg/tools/response_waiter.go b/pkg/tools/response_waiter.go
new file mode 100644
index 000000000..8ab8dbf3b
--- /dev/null
+++ b/pkg/tools/response_waiter.go
@@ -0,0 +1,52 @@
+package tools
+
+import "sync"
+
+// DeviceResponseWaiter is a package-level shared instance used to synchronize
+// tool_request/tool_response pairs between the Android tool and WebSocket channel.
+var DeviceResponseWaiter = NewResponseWaiter()
+
+// ResponseWaiter manages pending request/response synchronization.
+// Each request registers a channel by ID; the response is delivered when it arrives.
+type ResponseWaiter struct {
+	pending map[string]chan string
+	mu      sync.Mutex
+}
+
+func NewResponseWaiter() *ResponseWaiter {
+	return &ResponseWaiter{
+		pending: make(map[string]chan string),
+	}
+}
+
+// Register creates a buffered channel for the given request ID and returns it.
+// The caller should select on this channel with a timeout.
+func (w *ResponseWaiter) Register(id string) chan string {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	ch := make(chan string, 1)
+	w.pending[id] = ch
+	return ch
+}
+
+// Deliver sends the response content to the waiting channel for the given ID.
+// If no waiter is registered for the ID, the delivery is silently dropped.
+func (w *ResponseWaiter) Deliver(id, content string) {
+	w.mu.Lock()
+	ch, ok := w.pending[id]
+	if ok {
+		delete(w.pending, id)
+	}
+	w.mu.Unlock()
+	if ok {
+		ch <- content
+	}
+}
+
+// Cleanup removes the pending channel for the given ID without delivering.
+// Used on timeout to prevent memory leaks.
+func (w *ResponseWaiter) Cleanup(id string) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	delete(w.pending, id)
+}