An agent that treats your phone as a body, not a tool

What it does

Andee is a single Kotlin APK that takes over an Android phone (11+) and operates it the way you would: reading the accessibility tree, tapping, typing, handling notifications, SMS, calls, calendar, camera, even a robot dog. It runs its brain on-device against any OpenAI-compatible endpoint you point it at — OpenAI or DeepSeek by default, no vendor account. The UI is a floating ball that changes expression mid-sentence while it works.

The interesting bit

The pitch is less “assistant” and more “resident”: it keeps a local-only notebook of habits and promises, and after a conversation goes quiet it sends itself a (to yourself …) turn to reflect alone — biased toward silence, ending empty-handed rather than inventing things to record. Also notable: it never does coordinate math. Screenshots carry a labelled 8×12 grid and it taps by cell plus which third, which is exactly the crutch general models need.

Key highlights

• Works in any app with zero integrations — no SDK, no intent exposure; falls back to screenshots for WebViews and games.
• Types through the accessibility input channel on Android 13+, so your own keyboard stays active.
• Safety rules live in the dispatcher, not the prompt: it stops before final purchase confirmation, never enters payment passwords, and a per-location guard refuses the fourth tap on the same spot.
• Vault uses Keystore AES-GCM; the brain can type your password but no tool will ever return it, even masked.
• No analytics SDK, no content telemetry — the only outbound data is a version number and phone model on update checks.

Caveats

• The README is upfront that this is experimental and unwarrantied: it can complete payments, send messages under your name, and read contacts and location. What it does is your responsibility.
• Data does leave the device: speech goes to your chosen speech provider (Volcengine by default) and the conversation goes to whichever model you configure — “on-device” applies to the agent loop, not the LLM.
• Android only; the tablet/robot-dog ambitions are plans, not shipped code.

Verdict

Worth a look if you build mobile agents or care about the accessibility-tree-plus-grid approach to screen acting — the dispatcher-level guard rails are genuinely instructive. Skip it if you want a polished assistant or anything with a warranty-shaped object near it.