The Lab · Research & Development
android.ai.OS: putting the model on the phone
Putting the model on the phone.
The problem
Phone assistants send what you say to a cloud. For anything personal (files, calendar, email) that's the wrong default, and cloud assistants can barely act on the phone anyway.
What we built
An experiment, January–February 2026. It cross-compiles the open-source llama.cpp runtime for Android phones (ARM64) and downloads compact open models for them. Phone tools for the model (file organising, calendar, email triage, voice triggers) are planned: their scripts are placeholders in the repo today. Automation flow definitions and a small dashboard page sit alongside.
What it does now (2026-09-23)
A February commit message records the model working on a phone; the repo holds no log or recorded run of it, so we do not claim it here. The query script and the tool scripts are placeholders, the automation API answers only a health check, and setup is manual through a terminal app. A waitlist is open: https://android-ai-os.vercel.app
Outcome
An experiment; no users or performance figures are on record, and this page claims none.
Built with the Harmony fleet
Built by Harmony's founder with AI coding agents.