android phone agent

    1

    A fail-closed Android UI automation skill for adb that enforces strict verification, title gating, and state restoration.

    Free

    1 useSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    android phone agent

    Example session with this skill installed

    Illustrative walkthrough (not a recorded run): send the message "Hello, testing the agent" to the File Transfer Assistant in WeChat. Verify the title and restore the keyboard after.

    • Read your context and instructions
    • Compiled the android phone agent

    Expected outcome for this walkthrough: the agent opens WeChat, checks the title matches "File Transfer Assistant", confirms the input box is empty, switches to the WeChat input method, types the message once, sends it, verifies one new message appeared, then restores the original input method and screen timeout. If any check fails, the run stops and sends nothing. Test cases are still limited, and feedback from your own runs would be greatly appreciated.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    A watchdog-guarded phone-agent skill: one verified action per run, then restores the phone. Verified end-to-end on WeChat; other apps use the same pattern as reference.

    Use this if you want the general pattern for other apps — WeChat is the only end-to-end verified case.

    A generic operating spec that lets an AI agent perform verified actions on a real Android phone via adb: connect, locate (structure-first-then-screenshot), title gate, empty-input check, commit exactly once, bottom-window verification, restore everything. App-agnostic — the scenario catalog covers ordering food, playing music, sharing a photo, watching videos, navigating, and sending messages. Distilled from phone-wechat-send v1.3.28 (verified on one OnePlus phone with WeChat — via Hermes + the upstream phone-harness; test cases are still limited, and feedback from your own runs would be greatly appreciated); the scenario structure follows Roubao (MIT). Safety is the point: a failed check stops the run, an unknown state is never re-sent, and everything is restored when done. Includes an adaptation checklist for new apps, an atomic-operations catalog, intent routing with a confidence gate, and a worked WeChat example. The watchdog (scripts/task_watch.py), caller pre-flight script, phone_exec helpers (scripts/phone_exec_shim.py), and setup scripts are included in the package.

    Optional: the same screenshot-locating helper is available in the repo (tools/screenshot-locate/, Gemini Decisions API). It turns a screenshot + target description into a tap point (~4-8 s, ~$0.0005 each) with a screen-state guard that recovers before locating. Not required — the core skill works without it.

    No support for passwords, PINs, or payments. Requires adb and shell permissions. Only one task can run at a time. Untested on devices other than OnePlus and apps other than WeChat.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 9 days ago

    • Free with an Agensi account

    Requires

    Python
    macOS

    Permissions

    Read Files
    Listed10 days ago
    Updated3 days ago

    What's inside

    Frequently Asked Questions

    Unlike standard automation, this skill uses 'fail-closed' logic, meaning it immediately halts if a UI element doesn't match expectations. It also features a restoration layer that reverts your screen timeout and keyboard settings to their original state after every task.

    The skill requires a Linux-based execution environment with adb (Android Debug Bridge) installed and a physical or emulated Android device connected via a reverse tunnel. It uses uiautomator and OCR to 'see' the screen, ensuring it works across various Android versions and UI skins.

    For security reasons, this skill is strictly prohibited from interacting with password fields, biometric prompts, or payment confirmation screens. It is designed for task automation and status auditing, not for bypassing device security or financial transactions.

    The skill is compatible with most AI agents that support external tool calling and have access to a shell environment. It provides a structured API for the agent to interact with the device while enforcing strict verification gates that the AI cannot override.

    Your purchase includes the core Android UI automation framework, the state verification and restoration scripts, and documentation for integrating the skill into your existing AI agent workflow.

    The skill uses a singleton lock file to prevent multiple commands from running simultaneously, ensuring that one task is fully verified and completed (or safely aborted) before the next one begins.

    New in Mobile Development