Simple Agent

A minimal on-device LLM agent app for Android, built as a simple reference for consuming localllmhost — the ArkLlm host that runs a LiteRT-LM model on your phone.

This app does not run the model itself. It binds to the ArkLlm host over AIDL, sends your conversation, and streams back answers — exactly like any other app would use the host's client SDK. The "agent" part (system prompt, tool calling, local tool execution) lives here in the client, so this is a working example of the full flow: chat UI → agent loop → host inference → tools.

Features

• Jetpack Compose chat UI with streaming replies and a "thinking…" indicator
• Two tools exposed to the on-device model:
  • get_system_time — returns the device's current time/date
  • web_search — real web search via the Tavily API
• Agentic tool-calling loop: the model requests a tool, this app executes it locally and feeds the result back, repeating until a plain answer arrives
• Multi-turn context (full conversation history is replayed per session)
• Fully offline inference — the model runs on-device in the host app

How it works

┌────────────────────────┐   bind (signature-level)   ┌─────────────────────────┐
│   Simple Agent (this)  │ ─────────────────────────▶ │  ArkLlm Host app        │
│   AgentRunner loop     │  AIDL: ILlmService          │  LiteRT-LM engine       │
│    tools (time/search) │ ◀───────────────────────── │  session replay +       │
└────────────────────────┘   streaming callbacks       │  on-device generation   │
                                                       └─────────────────────────┘

• The host app (com.arkj.llmserver) loads and owns a single warm LiteRT-LM engine. Binding is gated by the BIND_LLM_SERVICE signature-level permission, so both apps must be signed with the same key.
• On each send, this app opens a host session (llm-host-client) and sends the whole conversation history; the host replays and streams tokens back.
• When the model emits a tool call, the stream stops, this app executes the tool (no network needed for get_system_time, Tavily API for web_search) and sends the result back for the next turn — up to 8 turns.

Repositories

Repo	Role
Ark946/localllmhost	The host app + llm-contract + llm-host-client SDK (LiteRT-LM backend)
this repo	A client app demonstrating how to use the host client

Prerequisites

1. Android Studio (or a gradlew-capable environment), AGP 9.x, JDK 17.
2. Build & install the ArkLlm host app from localllmhost, then in its UI select and download a model (the default CPU backend works out of the box).
3. Sign both apps with the same signing key — binding is signature-scoped.
4. (Optional) A Tavily API key to use web_search.

Build & run

# 1. (optional) add your Tavily key — never commit local.properties
#    app/local.properties:
#    tavilyApiKey=tvly-xxxxxxxxxxxxxxxxxxxx

./gradlew :app:assembleDebug

or open the project in Android Studio and press Run. Install on a device (or emulator) that already has the host app installed and a model selected.

Note: the Tavily key is compiled into the APK via BuildConfig. Fine for a personal app; if you distribute it, route search through a server instead.

Dependencies

Dependency	Version	Purpose
com.github.Ark946.localllmhost:llm-host-client	v0.1.0	Host client SDK (via JitPack)
dev.langchain4j:langchain4j-core	1.12.2	Tool-call types / tool spec schema
com.squareup.okhttp3:okhttp	4.12.0	Web search HTTP
com.google.code.gson:gson	2.13.2	JSON (tools, Tavily)
Jetpack Compose (BOM)	2026.02.01	UI

Minimum SDK 28, target SDK 36. Kotlin 2.2.10, AGP 9.1.1.

Project layout

app/src/main/java/com/arkj/simpleagent/
├── MainActivity.kt        # Compose chat UI
├── AgentTools.kt          # Tool specs, tool execution, and the agent loop
└── ui/theme/              # Material 3 theme

The interesting part for your own agent is AgentTools.kt: it defines the two tools, executes them, and runs the streaming tool-calling loop over RemoteLlmClient from llm-host-client.

License

Apache-2.0 — the same license as localllmhost.