📱 Hello World — On-Device Android

The model runs INSIDE the app. Airplane mode. In the video, qwen3-4b lives in the Ghost APK via llama.cpp — the agent opens Reddit, reads the frontpage, and summarizes it, and nothing ever leaves the phone. The phone is both the body and the brain.

Prerequisites

An Android phone with enough RAM for your chosen model (~4 GB free for a 4B Q4 model, less for Gemma 1B/2B)
The Ghost companion APK installed on the phone (built from the repo — see On-Device LLM)
A model file: .gguf (llama.cpp) or .task (MediaPipe) — the registry lists tested ones

Step 1 — Load a model

Open the Ghost app on the phone and pick a model from the on-device registry (e.g. qwen3-4b GGUF, or gemma-3-1b-it for smaller phones). OnDeviceLLM.ensureLoaded(modelId) routes to the right engine — MediaPipe for .task, llama.cpp (JNI) for .gguf. You don’t have to care which.

Step 2 — Flip on airplane mode

Seriously — do it. That’s the point. The demo runs with zero connectivity.

Step 3 — Ask in plain English, on the phone

In the Ghost app’s chat:

summarize the reddit frontpage

The agent launches Reddit, reads the screen tree, scrolls, comes back, and streams a summary — reasoned entirely on-device. The on-device · qwen3-4b chip in the UI confirms no cloud is involved.

Why this is different

	Where model runs	What leaves the phone
Cloud (Anthropic, OpenRouter…)	Cloud GPU	Prompts + tools
Local (Ollama, vLLM)	Your laptop/desk GPU	Prompts leave the phone, stay on your network
On-device ⭐	The phone itself	Nothing

What next?

On-Device LLM — engines, model registry, the full stack
Stealth Mode — no root, no cloud, no trace
LLM Providers — mix on-device with cloud brains per task

Try another approach →

MCP + Claude Code · MCP + Codex · MCP + Antigravity · Ghost CLI · LangChain · Web Chat · On-Device iPhone · iPhone Duolingo