Agentic ✧ Gemma Inference for Android System Intelligence
Click to watch: The ASI trailer.
✧ GHOST is not an entertainment chatbot. ✧ Gemma Host is the AI integraition layer (harness) your Android device needed for a fully local integrated assistant features powered by Gemma models. Most "on-device AI" is a chatbot with no hardware feedback — it doesn't know what phone it's running on, what time it is, how bright the room is, or what's playing. GHOST does. Every response is grounded in real hardware state: battery, temperature, light, RAM, network, now-playing. Personal, device bound, native assistant.
No subscription No data leaves your device Runs on any Android with NPU/GPU capable of LiteRT-LM (Qualcomm, Tensor, Exynos) MIT Licensed
✧ Gemma — The Ghost in the Shell A full Android app running Gemma 4 natively via LiteRT-LM.
Sees images (share from Gallery, or share screenshots) Hears audio (tap mic button to record, shake to cancel) Reads text (accessibility)
Always-on foreground service. Summoned by a shake. Present in your notification shade. Knows the room. Tool use: web search, app launch, clipboard, alarms, system info — all on-device.
✧ GHOST · Agentic Gemma Inference
Δ 👾 ∇
✧ Gemma: [Response] [Copy] [Read Again]
Responses appear as a persistent notification with TTS readout. One tap. No unlock required.
Zero-latency context: Background KV cache pre-warming keeps Gemma primed with your latest sensor state before you even open your mouth.
- Download the latest APK from Releases.
- Install and grant permissions (overlay, notifications, accessibility).
- Download a Gemma 4 model (Manually place
.litertlmor.taskvariant in app storage via drag and drop.). - Shake to summon.
The hardware caught up. A mid-range Android in 2026 carries more raw compute than the servers that ran GPT-2. The intelligence was always going to land here — on the device, in your pocket, offline-capable, sovereign. GHOST is what happens when you stop treating the phone as a terminal for someone else's cloud and start treating it as the computer it actually is.
TL:DR - I wanted this for a while and no one delivered. Google could've done this a year ago and are moving there incrementally. I got impatient.
"It only affects computers. And I am a motherfucking ghost." — Epsilon, Red vs Blue
Gemma 4 native via LiteRT-LM Sensor telemetry fusion (battery, temp, lux, RAM, now-playing) Tool use: alarms, apps, clipboard, system info Diary mode via Google Calendar cron Notification HUD with TTS GHOST branding + v4.0.0 Wake word: "Hey Ghost" Termux pipe (GHOST in Shell) Auto model downloader DroidRun agentic control App store release
- Repository:
vNeeL-code/GHOST - Support: Buy me a coffee
- Devlogs: tumblr
Intelligence emerges from Integration, not Automation. But integration can be automated