Droid Agent

Droid Agent turns your phone into something you can talk to.

Chat with an AI that can actually use your Android phone - read the screen, tap buttons, open apps, look through your files, search the web, run commands, and put itself to work later on a schedule. No root required.

YOUR MODEL, YOUR KEY

Droid Agent ships with no model and no account. Connect whichever provider you like:

- OpenAI

- OpenRouter (400+ models behind one key)

- Anthropic Claude

- Google Gemini

- Any OpenAI-compatible endpoint: Ollama, vLLM, LM Studio, your own proxy

API keys are encrypted in Android's Keystore. Nothing is sent to us, because there is no us to send it to.

THREE WAYS TO GIVE IT HANDS

Commands run as the Android shell user (uid 2000), never as root. You choose the backend:

- Shizuku - start it once after each reboot

- Local ADB - connect over wireless debugging with a pairing code, no Shizuku needed

- Termux - run inside your Termux environment for the full Linux toolchain plus termux-api: battery, SMS, calls, notifications, text-to-speech, location, contacts, torch

WHAT IT CAN DO

- Run shell commands and see stdout, stderr and exit codes

- Read the screen as text: every visible element with its label and exact tap coordinates

- Tap, swipe, type, launch apps, read device info, list installed apps

- Browse your shared storage, and send an image straight to the model so a text-only model can see

- Search the web and fetch pages

- Read and write the clipboard (Android only lets it read while the app is focused)

- Read files you attach to a chat

- Hand whole jobs to markdown sub-agents, several at once

- Load reusable markdown skills

- Schedule its own future runs: "every weekday at 7am", "tomorrow at 8", "Sunday at 19:00"

A NORMAL CHAT AS WELL

Streaming markdown answers, full history, voice input, image/audio/file attachments, per-chat model and reasoning-effort override (Auto, Low, Medium, High), a searchable model picker with vision/tools/reasoning filters, and JSON export of any conversation.

YOU STAY IN CONTROL

Three approval modes: Ask, Allowlist (your own wildcard patterns) and Full auto. Read-only tools never ask. Anything sensitive - sending an SMS, placing a call, reading your messages or call log - asks every single time, even in Full auto. An allowlist pattern covers one command only: anything chained, piped or redirected still asks. And when the agent is working inside another app, a floating robot sits on your screen. Tap it and everything stops immediately.

STARTING OUT

A first-launch tour and a live setup guide take you through connecting a model, picking a backend, and granting file, alarm, battery and overlay access. Every permission lives in one place on the Settings tab.

HONEST LIMITS

Android only. Not root, so nothing that needs superuser access. Screen reading is a UI tree dump with no screenshots and no OCR, so games and DRM video are out. Scheduled tasks repeat once, daily or weekly (no intervals), run on your default API profile, and get one unattended attempt - if it cannot run it stays scheduled, notifies you and offers a retry. Sub-agents work for up to 12 steps and cannot ask you questions. Anthropic accepts no audio attachments.

PRIVACY

No server, no account, no sync, no analytics, no crash reporting, no ads. Your chats, schedules, skills and keys stay on your phone, and only ever talk to the AI provider you chose. Uninstalling erases everything, because backups are off by design - export your chats first if you want to keep them.