feat: Add Toggle Server for local inference

This commit introduces a new "Toggle Server" feature that runs a local HTTP server on the device. This allows developers and researchers to interact with the on-device AI models using `curl`, with all communication tunneled exclusively over the USB cable.

The server can handle multipart/form-data requests, allowing users to send a prompt, an image, or both. This provides a powerful new way to test, debug, and integrate the on-device models.

.gitignore

@@ -1 +1,3 @@
 .DS_Store
+.idea/
+.gemini/

Android/src/app/build.gradle.kts

@@ -99,6 +99,7 @@ dependencies {
   implementation(libs.hilt.navigation.compose)
   implementation(platform(libs.firebase.bom))
   implementation(libs.firebase.analytics)
+  implementation("commons-fileupload:commons-fileupload:1.4")
   kapt(libs.hilt.android.compiler)
   testImplementation(libs.junit)
   androidTestImplementation(libs.androidx.junit)
@@ -113,4 +114,4 @@ dependencies {
 protobuf {
   protoc { artifact = "com.google.protobuf:protoc:4.26.1" }
   generateProtoTasks { all().forEach { it.plugins { create("java") { option("lite") } } } }
-}
+}

Android/src/app/src/main/AndroidManifest.xml

@@ -32,6 +32,7 @@
     <uses-permission android:name="android.permission.RECORD_AUDIO" />
     <uses-permission android:name="android.permission.WAKE_LOCK"/>
     <uses-permission android:name="android.permission.ACCESS_NETWORK_STATE" />
+    <uses-permission android:name="android.permission.READ_EXTERNAL_STORAGE" />
 
     <uses-feature
         android:name="android.hardware.camera"
@@ -47,6 +48,7 @@
         android:roundIcon="@mipmap/ic_launcher"
         android:supportsRtl="true"
         android:theme="@style/Theme.Gallery"
+        android:enableOnBackInvokedCallback="true"
         tools:targetApi="31">
         <activity
             android:name="com.google.ai.edge.gallery.MainActivity"

Android/src/app/src/main/java/com/google/ai/edge/gallery/MainActivity.kt

@@ -26,6 +26,7 @@ import androidx.activity.enableEdgeToEdge
 import androidx.compose.foundation.layout.fillMaxSize
 import androidx.compose.material3.Surface
 import androidx.compose.ui.Modifier
+import androidx.core.app.ActivityCompat
 import androidx.core.splashscreen.SplashScreen.Companion.installSplashScreen
 import com.google.ai.edge.gallery.ui.theme.GalleryTheme
 import com.google.firebase.analytics.FirebaseAnalytics
@@ -60,6 +61,10 @@ class MainActivity : ComponentActivity() {
     setContent { GalleryTheme { Surface(modifier = Modifier.fillMaxSize()) { GalleryApp() } } }
     // Keep the screen on while the app is running for better demo experience.
     window.addFlags(WindowManager.LayoutParams.FLAG_KEEP_SCREEN_ON)
+
+    if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.M) {
+      requestPermissions(arrayOf(android.Manifest.permission.READ_EXTERNAL_STORAGE), 1)
+    }
   }
 
   companion object {

Android/src/app/src/main/java/com/google/ai/edge/gallery/data/Tasks.kt

@@ -33,6 +33,7 @@ enum class TaskType(val label: String, val id: String) {
   LLM_PROMPT_LAB(label = "Prompt Lab", id = "llm_prompt_lab"),
   LLM_ASK_IMAGE(label = "Ask Image", id = "llm_ask_image"),
   LLM_ASK_AUDIO(label = "Audio Scribe", id = "llm_ask_audio"),
+  TOGGLE_SERVER(label = "Toggle Server", id = "toggle_server"),
   TEST_TASK_1(label = "Test task 1", id = "test_task_1"),
   TEST_TASK_2(label = "Test task 2", id = "test_task_2"),
 }
@@ -121,9 +122,20 @@ val TASK_LLM_ASK_AUDIO =
     textInputPlaceHolderRes = R.string.text_input_placeholder_llm_chat,
   )
 
+val TASK_TOGGLE_SERVER =
+  Task(
+    type = TaskType.TOGGLE_SERVER,
+    icon = Icons.Outlined.Forum,
+    models = mutableListOf(),
+    description = "Toggle an LLM endpoint server running on-device (Placeholder).",
+    docUrl = "",
+    sourceCodeUrl = "",
+    textInputPlaceHolderRes = R.string.text_input_placeholder_llm_chat,
+  )
+
 /** All tasks. */
 val TASKS: List<Task> =
-  listOf(TASK_LLM_ASK_IMAGE, TASK_LLM_ASK_AUDIO, TASK_LLM_PROMPT_LAB, TASK_LLM_CHAT)
+  listOf(TASK_LLM_ASK_IMAGE, TASK_LLM_ASK_AUDIO, TASK_LLM_PROMPT_LAB, TASK_LLM_CHAT, TASK_TOGGLE_SERVER)
 
 fun getModelByName(name: String): Model? {
   for (task in TASKS) {

Android/src/app/src/main/java/com/google/ai/edge/gallery/server/InAppServer.kt

@@ -0,0 +1,278 @@
+package com.google.ai.edge.gallery.server
+
+import android.content.Context
+import android.graphics.Bitmap
+import android.graphics.BitmapFactory
+import android.util.Log
+import com.google.ai.edge.gallery.data.TASK_LLM_ASK_IMAGE
+import com.google.ai.edge.gallery.ui.llmchat.LlmChatModelHelper
+import dagger.hilt.android.qualifiers.ApplicationContext
+import java.io.ByteArrayInputStream
+import java.io.ByteArrayOutputStream
+import java.io.IOException
+import java.io.InputStream
+import java.io.PrintWriter
+import java.net.ServerSocket
+import java.net.Socket
+import java.net.SocketException
+import java.net.URLDecoder
+import java.util.concurrent.CountDownLatch
+import javax.inject.Inject
+import javax.inject.Singleton
+import org.apache.commons.fileupload.FileItem
+import org.apache.commons.fileupload.disk.DiskFileItemFactory
+import org.apache.commons.fileupload.servlet.ServletFileUpload
+
+@Singleton
+class InAppServer @Inject constructor(
+    @ApplicationContext private val context: Context,
+    private val llmChatModelHelper: LlmChatModelHelper
+) {
+
+    private var serverSocket: ServerSocket? = null
+    private var serverThread: Thread? = null
+    @Volatile
+    private var isServerRunning = false
+
+    fun start() {
+        if (isServerRunning) return
+
+        serverThread = Thread {
+            try {
+                llmChatModelHelper.initialize(context, TASK_LLM_ASK_IMAGE.models.first()) {
+                    if (it.isNotEmpty()) {
+                        Log.e(TAG, "Failed to initialize model: $it")
+                        return@initialize
+                    }
+                }
+
+                serverSocket = ServerSocket(DEVICE_PORT)
+                isServerRunning = true
+                Log.i(TAG, "In-App Server started on port " + DEVICE_PORT)
+
+                while (isServerRunning) {
+                    try {
+                        val clientSocket = serverSocket!!.accept()
+                        Log.i(TAG, "Client connected: " + clientSocket.inetAddress)
+                        handleClient(clientSocket)
+                    } catch (e: SocketException) {
+                        if (!isServerRunning) {
+                            Log.i(TAG, "Server socket closed intentionally.")
+                        } else {
+                            Log.e(TAG, "Error accepting connection", e)
+                        }
+                    }
+                }
+            } catch (e: IOException) {
+                Log.e(TAG, "Error starting server", e)
+                isServerRunning = false
+            }
+        }
+        serverThread!!.start()
+    }
+
+    fun stop() {
+        if (!isServerRunning) return
+
+        try {
+            isServerRunning = false
+            if (serverSocket != null && !serverSocket!!.isClosed) {
+                serverSocket!!.close()
+            }
+            if (serverThread != null) {
+                serverThread!!.interrupt()
+                serverThread = null
+            }
+            llmChatModelHelper.cleanUp(TASK_LLM_ASK_IMAGE.models.first())
+            Log.i(TAG, "In-App Server stopped.")
+        } catch (e: IOException) {
+            Log.e(TAG, "Error stopping server", e)
+        }
+    }
+
+    private fun handleClient(clientSocket: Socket) {
+        try {
+            val inputStream = clientSocket.inputStream
+            val writer = PrintWriter(clientSocket.outputStream, true)
+
+            val requestLine = readLine(inputStream)
+            if (requestLine.isBlank()) {
+                clientSocket.close()
+                return
+            }
+            Log.i(TAG, "Request: $requestLine")
+
+            val requestParts = requestLine.split(" ")
+            val method = requestParts[0]
+
+            var contentType = ""
+            var contentLength = 0
+            var line = readLine(inputStream)
+            while (line.isNotEmpty()) {
+                if (line.startsWith("Content-Type:", ignoreCase = true)) {
+                    contentType = line.substringAfter(":").trim()
+                } else if (line.startsWith("Content-Length:", ignoreCase = true)) {
+                    contentLength = line.substringAfter(":").trim().toInt()
+                }
+                line = readLine(inputStream)
+            }
+
+            var prompt = ""
+            var imageData: ByteArray? = null
+
+            if (method == "POST") {
+                if (contentLength > 0) {
+                    val bodyBytes = ByteArray(contentLength)
+                    var bytesRead = 0
+                    while (bytesRead < contentLength) {
+                        val read = inputStream.read(bodyBytes, bytesRead, contentLength - bytesRead)
+                        if (read == -1) break
+                        bytesRead += read
+                    }
+
+                    if (ServletFileUpload.isMultipartContent(RequestContext(ByteArrayInputStream(bodyBytes), contentType, contentLength))) {
+                        val factory = DiskFileItemFactory()
+                        val upload = ServletFileUpload(factory)
+                        val items = upload.parseRequest(RequestContext(ByteArrayInputStream(bodyBytes), contentType, contentLength))
+                        for (item in items) {
+                            if (item.isFormField) {
+                                if (item.fieldName == "prompt") {
+                                    prompt = item.string
+                                }
+                            } else {
+                                if (item.fieldName == "image") {
+                                    imageData = item.get()
+                                }
+                            }
+                        }
+                    } else {
+                        prompt = String(bodyBytes)
+                    }
+                }
+            } else { // GET
+                val queryParams = getQueryParams(requestLine)
+                prompt = queryParams["prompt"] ?: ""
+            }
+
+            if (prompt.isBlank() && imageData == null) {
+                writer.println("HTTP/1.1 400 Bad Request")
+                writer.println("Content-Type: text/plain")
+                writer.println()
+                writer.println("No prompt or image provided.")
+                writer.flush()
+                clientSocket.close()
+                return
+            }
+
+            writer.println("HTTP/1.1 200 OK")
+            writer.println("Content-Type: text/plain")
+            writer.println("Connection: close")
+            writer.println()
+            writer.flush()
+
+            val latch = CountDownLatch(1)
+            llmChatModelHelper.resetSession(TASK_LLM_ASK_IMAGE.models.first())
+
+            val images: List<Bitmap> = imageData?.let {
+                val bitmap = BitmapFactory.decodeByteArray(it, 0, it.size)
+                listOf(bitmap)
+            } ?: emptyList()
+
+            llmChatModelHelper.runInference(
+                model = TASK_LLM_ASK_IMAGE.models.first(),
+                input = prompt,
+                images = images,
+                resultListener = { partialResult, done ->
+                    writer.print(partialResult)
+                    writer.flush()
+                    if (done) {
+                        clientSocket.close()
+                        latch.countDown()
+                    }
+                },
+                cleanUpListener = {
+                    if (!clientSocket.isClosed) {
+                        writer.flush()
+                        clientSocket.close()
+                    }
+                    latch.countDown()
+                }
+            )
+            latch.await()
+        } catch (e: Exception) {
+            Log.e(TAG, "Error handling client", e)
+        } finally {
+            try {
+                if (!clientSocket.isClosed) {
+                    clientSocket.close()
+                }
+            } catch (e: IOException) {
+                Log.e(TAG, "Error closing client socket", e)
+            }
+        }
+    }
+
+    private fun readLine(stream: InputStream): String {
+        val buffer = ByteArrayOutputStream()
+        while (true) {
+            val b = stream.read()
+            if (b == -1) break
+            if (b == '\n'.code) {
+                break
+            }
+            buffer.write(b)
+        }
+        val bytes = buffer.toByteArray()
+        if (bytes.isNotEmpty() && bytes.last() == '\r'.toByte()) {
+            return String(bytes, 0, bytes.size - 1, Charsets.ISO_8859_1)
+        }
+        return String(bytes, Charsets.ISO_8859_1)
+    }
+
+    private fun getQueryParams(requestLine: String): Map<String, String> {
+        val queryParams = mutableMapOf<String, String>()
+        val urlParts = requestLine.split(" ")[1].split("?")
+        if (urlParts.size > 1) {
+            val query = urlParts[1]
+            for (param in query.split("&")) {
+                val pair = param.split("=")
+                if (pair.size > 1) {
+                    queryParams[URLDecoder.decode(pair[0], "UTF-8")] =
+                        URLDecoder.decode(pair[1], "UTF-8")
+                }
+            }
+        }
+        return queryParams
+    }
+
+    fun isRunning(): Boolean {
+        return isServerRunning
+    }
+
+    companion object {
+        private const val TAG = "AIEdgeServer"
+        private const val DEVICE_PORT = 8080
+    }
+}
+
+class RequestContext(
+    private val inputStream: java.io.InputStream,
+    private val contentType: String,
+    private val contentLength: Int
+) : org.apache.commons.fileupload.RequestContext {
+    override fun getCharacterEncoding(): String {
+        return "UTF-8"
+    }
+
+    override fun getContentType(): String {
+        return contentType
+    }
+
+    override fun getContentLength(): Int {
+        return contentLength
+    }
+
+    override fun getInputStream(): java.io.InputStream {
+        return inputStream
+    }
+}

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatModelHelper.kt

@@ -32,6 +32,8 @@ import com.google.mediapipe.framework.image.BitmapImageBuilder
 import com.google.mediapipe.tasks.genai.llminference.GraphOptions
 import com.google.mediapipe.tasks.genai.llminference.LlmInference
 import com.google.mediapipe.tasks.genai.llminference.LlmInferenceSession
+import javax.inject.Inject
+import javax.inject.Singleton
 
 private const val TAG = "AGLlmChatModelHelper"
 
@@ -41,7 +43,8 @@ typealias CleanUpListener = () -> Unit
 
 data class LlmModelInstance(val engine: LlmInference, var session: LlmInferenceSession)
 
-object LlmChatModelHelper {
+@Singleton
+class LlmChatModelHelper @Inject constructor() {
   // Indexed by model name.
   private val cleanUpListeners: MutableMap<String, CleanUpListener> = mutableMapOf()
 

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmchat/LlmChatViewModel.kt

@@ -51,7 +51,10 @@ private val STATS =
     Stat(id = "latency", label = "Latency", unit = "sec"),
   )
 
-open class LlmChatViewModelBase(val curTask: Task) : ChatViewModel(task = curTask) {
+open class LlmChatViewModelBase(
+    val curTask: Task,
+    private val llmChatModelHelper: LlmChatModelHelper
+) : ChatViewModel(task = curTask) {
   fun generateResponse(
     model: Model,
     input: String,
@@ -92,7 +95,7 @@ open class LlmChatViewModelBase(val curTask: Task) : ChatViewModel(task = curTas
       val start = System.currentTimeMillis()
 
       try {
-        LlmChatModelHelper.runInference(
+        llmChatModelHelper.runInference(
           model = model,
           input = input,
           images = images,
@@ -195,7 +198,7 @@ open class LlmChatViewModelBase(val curTask: Task) : ChatViewModel(task = curTas
 
       while (true) {
         try {
-          LlmChatModelHelper.resetSession(model = model)
+          llmChatModelHelper.resetSession(model = model)
           break
         } catch (e: Exception) {
           Log.d(TAG, "Failed to reset session. Trying again")
@@ -262,12 +265,16 @@ open class LlmChatViewModelBase(val curTask: Task) : ChatViewModel(task = curTas
 }
 
 @HiltViewModel
-class LlmChatViewModel @Inject constructor() : LlmChatViewModelBase(curTask = TASK_LLM_CHAT)
+class LlmChatViewModel @Inject constructor(
+    llmChatModelHelper: LlmChatModelHelper
+) : LlmChatViewModelBase(curTask = TASK_LLM_CHAT, llmChatModelHelper = llmChatModelHelper)
 
 @HiltViewModel
-class LlmAskImageViewModel @Inject constructor() :
+class LlmAskImageViewModel @Inject constructor(
-  LlmChatViewModelBase(curTask = TASK_LLM_ASK_IMAGE)
+    llmChatModelHelper: LlmChatModelHelper
+) : LlmChatViewModelBase(curTask = TASK_LLM_ASK_IMAGE, llmChatModelHelper = llmChatModelHelper)
 
 @HiltViewModel
-class LlmAskAudioViewModel @Inject constructor() :
+class LlmAskAudioViewModel @Inject constructor(
-  LlmChatViewModelBase(curTask = TASK_LLM_ASK_AUDIO)
+    llmChatModelHelper: LlmChatModelHelper
+) : LlmChatViewModelBase(curTask = TASK_LLM_ASK_AUDIO, llmChatModelHelper = llmChatModelHelper)

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/llmsingleturn/LlmSingleTurnViewModel.kt

@@ -66,7 +66,9 @@ private val STATS =
   )
 
 @HiltViewModel
-class LlmSingleTurnViewModel @Inject constructor() : ViewModel() {
+class LlmSingleTurnViewModel @Inject constructor(
+    private val llmChatModelHelper: LlmChatModelHelper
+) : ViewModel() {
   private val _uiState = MutableStateFlow(createUiState(task = TASK_LLM_PROMPT_LAB))
   val uiState = _uiState.asStateFlow()
 
@@ -80,7 +82,7 @@ class LlmSingleTurnViewModel @Inject constructor() : ViewModel() {
         delay(100)
       }
 
-      LlmChatModelHelper.resetSession(model = model)
+      llmChatModelHelper.resetSession(model = model)
       delay(500)
 
       // Run inference.
@@ -96,7 +98,7 @@ class LlmSingleTurnViewModel @Inject constructor() : ViewModel() {
       val start = System.currentTimeMillis()
       var response = ""
       var lastBenchmarkUpdateTs = 0L
-      LlmChatModelHelper.runInference(
+      llmChatModelHelper.runInference(
         model = model,
         input = input,
         resultListener = { partialResult, done ->

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/modelmanager/ModelManagerViewModel.kt

@@ -144,6 +144,7 @@ constructor(
   private val downloadRepository: DownloadRepository,
   private val dataStoreRepository: DataStoreRepository,
   private val lifecycleProvider: AppLifecycleProvider,
+  private val llmChatModelHelper: LlmChatModelHelper,
   @ApplicationContext private val context: Context,
 ) : ViewModel() {
   private val externalFilesDir = context.getExternalFilesDir(null)
@@ -292,8 +293,8 @@ constructor(
         TaskType.LLM_ASK_IMAGE,
         TaskType.LLM_ASK_AUDIO,
         TaskType.LLM_PROMPT_LAB ->
-          LlmChatModelHelper.initialize(context = context, model = model, onDone = onDone)
+          llmChatModelHelper.initialize(context = context, model = model, onDone = onDone)
-
+        TaskType.TOGGLE_SERVER -> {}
         TaskType.TEST_TASK_1 -> {}
         TaskType.TEST_TASK_2 -> {}
       }
@@ -308,8 +309,8 @@ constructor(
         TaskType.LLM_CHAT,
         TaskType.LLM_PROMPT_LAB,
         TaskType.LLM_ASK_IMAGE,
-        TaskType.LLM_ASK_AUDIO -> LlmChatModelHelper.cleanUp(model = model)
+        TaskType.LLM_ASK_AUDIO -> llmChatModelHelper.cleanUp(model = model)
-
+        TaskType.TOGGLE_SERVER -> {}
         TaskType.TEST_TASK_1 -> {}
         TaskType.TEST_TASK_2 -> {}
       }

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/navigation/Destination.kt

@@ -0,0 +1,5 @@
+package com.google.ai.edge.gallery.ui.navigation
+
+interface Destination {
+    val route: String
+}

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/navigation/GalleryNavGraph.kt

@@ -69,6 +69,8 @@ import com.google.ai.edge.gallery.ui.llmsingleturn.LlmSingleTurnScreen
 import com.google.ai.edge.gallery.ui.llmsingleturn.LlmSingleTurnViewModel
 import com.google.ai.edge.gallery.ui.modelmanager.ModelManager
 import com.google.ai.edge.gallery.ui.modelmanager.ModelManagerViewModel
+import com.google.ai.edge.gallery.ui.toggleserver.ToggleServerDestination
+import com.google.ai.edge.gallery.ui.toggleserver.ToggleServerScreen
 
 private const val TAG = "AGGalleryNavGraph"
 private const val ROUTE_PLACEHOLDER = "placeholder"
@@ -143,7 +145,14 @@ fun GalleryNavHost(
     modelManagerViewModel = modelManagerViewModel,
     navigateToTaskScreen = { task ->
       pickedTask = task
-      showModelManager = true
+      if (task.type == TaskType.TOGGLE_SERVER) {
+        navigateToTaskScreen(
+          navController = navController,
+          taskType = task.type,
+        )
+      } else {
+        showModelManager = true
+      }
     },
   )
 
@@ -260,6 +269,14 @@ fun GalleryNavHost(
         )
       }
     }
+    
+    composable(
+      route = ToggleServerDestination.route,
+      enterTransition = { slideEnter() },
+      exitTransition = { slideExit() },
+    ) {
+      ToggleServerScreen()
+    }
   }
 
   // Handle incoming intents for deep links
@@ -294,6 +311,7 @@ fun navigateToTaskScreen(
     TaskType.LLM_ASK_AUDIO -> navController.navigate("${LlmAskAudioDestination.route}/${modelName}")
     TaskType.LLM_PROMPT_LAB ->
       navController.navigate("${LlmSingleTurnDestination.route}/${modelName}")
+    TaskType.TOGGLE_SERVER -> navController.navigate(ToggleServerDestination.route)
     TaskType.TEST_TASK_1 -> {}
     TaskType.TEST_TASK_2 -> {}
   }

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/toggleserver/ToggleServerDestination.kt

@@ -0,0 +1,7 @@
+package com.google.ai.edge.gallery.ui.toggleserver
+
+import com.google.ai.edge.gallery.ui.navigation.Destination
+
+object ToggleServerDestination : Destination {
+    override val route = "toggle_server"
+}

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/toggleserver/ToggleServerScreen.kt

@@ -0,0 +1,30 @@
+package com.google.ai.edge.gallery.ui.toggleserver
+
+import androidx.compose.foundation.layout.Arrangement
+import androidx.compose.foundation.layout.Column
+import androidx.compose.foundation.layout.fillMaxSize
+import androidx.compose.material3.Button
+import androidx.compose.material3.Text
+import androidx.compose.runtime.Composable
+import androidx.compose.runtime.collectAsState
+import androidx.compose.runtime.getValue
+import androidx.compose.ui.Alignment
+import androidx.compose.ui.Modifier
+import androidx.hilt.navigation.compose.hiltViewModel
+
+@Composable
+fun ToggleServerScreen(
+    toggleServerViewModel: ToggleServerViewModel = hiltViewModel()
+) {
+    val isServerRunning by toggleServerViewModel.isServerRunning.collectAsState()
+
+    Column(
+        modifier = Modifier.fillMaxSize(),
+        verticalArrangement = Arrangement.Center,
+        horizontalAlignment = Alignment.CenterHorizontally
+    ) {
+        Button(onClick = { toggleServerViewModel.toggleServer() }) {
+            Text(if (isServerRunning) "Stop In-App Server" else "Start In-App Server")
+        }
+    }
+}

Android/src/app/src/main/java/com/google/ai/edge/gallery/ui/toggleserver/ToggleServerViewModel.kt

@@ -0,0 +1,33 @@
+package com.google.ai.edge.gallery.ui.toggleserver
+
+import androidx.lifecycle.ViewModel
+import com.google.ai.edge.gallery.server.InAppServer
+import dagger.hilt.android.lifecycle.HiltViewModel
+import javax.inject.Inject
+import kotlinx.coroutines.flow.MutableStateFlow
+import kotlinx.coroutines.flow.StateFlow
+import android.util.Log
+
+@HiltViewModel
+class ToggleServerViewModel @Inject constructor(
+    private val inAppServer: InAppServer
+) : ViewModel() {
+
+    private val _isServerRunning = MutableStateFlow(inAppServer.isRunning())
+    val isServerRunning: StateFlow<Boolean> = _isServerRunning
+
+    fun toggleServer() {
+        Log.d("ToggleServerViewModel", "toggleServer called")
+        if (inAppServer.isRunning()) {
+            inAppServer.stop()
+        } else {
+            inAppServer.start()
+        }
+        _isServerRunning.value = inAppServer.isRunning()
+    }
+
+    override fun onCleared() {
+        super.onCleared()
+        inAppServer.stop()
+    }
+}

README.md

@@ -19,6 +19,34 @@ The Google AI Edge Gallery is an experimental app that puts the power of cutting
 **AI Chat**
 <img width="1532" alt="AI Chat" src="https://github.com/user-attachments/assets/edaa4f89-237a-4b84-b647-b3c4631f09dc" />
 
+## 🔌 Toggle Server
+
+The "Toggle Server" feature runs a local HTTP server on your mobile device that allows you to interact with the on-device AI models from your laptop using `curl`, with all communication tunneled exclusively over a USB cable connection.
+
+### Usage
+
+1.  **Enable USB Debugging**:
+    *  Follow these [steps](https://developer.android.com/studio/debug/dev-options) to enable ADB port forwarding between your device and computer.
+
+2.  **Connect Device to Computer & Enable Port Forwarding**:
+    ```bash
+    adb -d forward tcp:8080 tcp:8080
+    ```
+
+3.  **Start the Server in the App**:
+    *   Navigate to the "Toggle Server" screen.
+    *   Tap the "Start In-App Server" button.
+
+4.  **Send Requests with `curl`**:
+    *   **Prompt only**:
+        ```bash
+        curl -X POST -F "prompt=Hello, world!" http://localhost:8080
+        ```
+    *   **Image and prompt**:
+        ```bash
+        curl -X POST -F "prompt=What is in this image?" -F "image=@/path/to/your/image.jpg" http://localhost:8080
+        ```
+
 ## ✨ Core Features
 
 *   **📱 Run Locally, Fully Offline:** Experience the magic of GenAI without an internet connection. All processing happens directly on your device.