Integrate EdgeLM in a few lines.

Pull the SDK from JitPack.

In settings.gradle.kts (or your root build.gradle), add the JitPack repo:

dependencyResolutionManagement {
    repositories {
        google()
        mavenCentral()
        maven { url = uri("https://jitpack.io") }
    }
}

Then in your app module's build.gradle.kts:

dependencies {
    implementation("com.github.Chandra-Mauli-Sharma.EdgeLM:sdk:0.1.0")
}

Check up front, prompt if missing.

EdgeLM is a shared runtime, so the EdgeLM Runtime app must be installed — the user gets it once from Play, and every EdgeLM app then shares it.

when (EdgeLM.status(context)) {
    EdgeLM.Status.NOT_INSTALLED -> EdgeLM.promptInstall(context)  // sends user to Play
    EdgeLM.Status.AVAILABLE     -> EdgeLM.initialize(context)     // bind the service
}

An OpenAI-style, coroutine-native API.

chat(...) returns a cold Flow<String> of tokens — work starts on collect, and cancelling the collector's scope cancels the on-device decode.

EdgeLM.initialize(context)          // once (e.g. Application.onCreate)

lifecycleScope.launch {
    try {
        EdgeLM.chat(
            model = "default",       // Phase 0 serves the active model; value is advisory
            prompt = "Explain on-device AI in one line",
            sessionId = "chat-1",    // stable id = multi-turn; "" = stateless one-shot
            priority = EdgeLM.INTERACTIVE,
        ).collect { token ->
            appendToUi(token)        // tokens stream in
        }
    } catch (e: EdgeLMUnavailableException) {
        // runtime missing or not responding — fall back / prompt install
    }
}

Multi-turn conversations

Pass a stable sessionId. Prior turns stay in the runtime's warm KV cache and aren't re-prefilled, so follow-ups are fast:

EdgeLM.chat("default", "My name is Ada.", sessionId = "s1").collect { … }
EdgeLM.chat("default", "What's my name?", sessionId = "s1").collect { … }  // remembers

Priority

Requests are admitted to the single shared engine by priority, with aging:

Three calls, start to finish.

Good to know.

• The user picks/downloads models in the EdgeLM Runtime app, not in yours. Your app just requests generations; whatever model is active serves them.
• Everything runs on-device — prompts and responses never leave the phone.
• If status() is NOT_INSTALLED, always give the user a path (promptInstall) rather than failing silently.