AI-powered invisible hand on your phone
GhostHand is an experimental on-device AI control system for Android.
It explores a simple but powerful idea:
an AI agent running inside an Android phone can control that same phone through a local execution environment and a wireless ADB bridge.
Instead of relying on a separate PC or remote server, GhostHand aims to turn the phone itself into a self-operating environment for execution, automation, prototyping, and device control.
GhostHand combines:
- Termux as the local Linux-like execution environment
- OpenClaw as the AI agent runtime
- Wireless ADB as the bridge into Android system control
- Natural language commands as the user interface
In practice, this means a user can say things like:
- "Open the clock app"
- "Launch Chrome and open Naver"
- "Turn on the flashlight"
- "Build and install this APK"
- "Open the dialer with this phone number"
…and GhostHand can translate those requests into real device actions.
Most device automation systems assume one of these models:
- a PC controls a phone
- a cloud agent controls an external device
- a macro tool automates UI interactions
GhostHand explores a different model:
- the AI agent runs on the phone itself
- the phone becomes both controller and target
- the same device can execute code, build apps, install them, and control its own Android layer
This makes GhostHand more than a simple automation script. It is closer to an on-device AI control layer for Android.
A simplified GhostHand flow looks like this:
User -> Chat / Natural Language Command
-> OpenClaw Agent in Termux
-> Command Planning / Interpretation
-> Wireless ADB Self-Connection
-> Android App / System / UI Control
-> Result Checking / Retry / Feedback
The user gives natural language instructions through chat or another lightweight interface.
OpenClaw interprets the request, decides what should happen, and chooses the control strategy.
Termux provides the local shell/runtime environment where scripts, commands, and build tools run.
Wireless ADB allows GhostHand to control the Android layer of the same device.
GhostHand can inspect UI state, package state, or command output to verify whether an action succeeded.
Current proof-of-concept capabilities include:
- Launch Android apps
- Open websites in Chrome
- Open settings and system panels
- Toggle quick settings items such as flashlight
- Open call screens and SMS compose screens
- Open messaging apps and specific chat targets when possible
- Build Android APKs directly from Termux
- Install and launch APKs on the same device
- Support small app prototyping workflows entirely on-device
GhostHand has already been used to test flows such as:
- opening Clock, Chrome, Gallery, Camera, YouTube, and Play Store
- searching content in Chrome and YouTube
- installing and launching a calculator app
- creating a simple Hello Android app in Termux
- building APKs directly on-device
- installing and running the built app
- creating and testing a minimal calculator app
- toggling flashlight from quick settings
- opening dialer and SMS compose flows
- opening Telegram chat targets through deep links
This means GhostHand is not just a concept note — it already has a working proof-of-concept base.
GhostHand sits at the intersection of:
- on-device AI agents
- Android system control
- self-hosted mobile automation
- local app development and deployment
- natural language interfaces
Its most interesting property is this:
the device is no longer just the thing being controlled — it also becomes the thing doing the controlling.
That opens the door to a new kind of personal mobile agent.
Status: early prototype / proof of concept
GhostHand is still experimental. It is currently focused on:
- validating the architecture
- improving reliability of same-device control
- documenting practical workflows
- building repeatable on-device automation patterns
Planned directions include:
- better command routing and fallback strategies
- reusable Android control modules
- more robust app-specific automations
- same-device AI orchestration workflows
- packaging GhostHand into a cleaner reproducible toolkit
- documenting patentable / research-worthy technical claims
GhostHand is guided by a few core principles:
- On-device first — avoid unnecessary external dependencies
- Natural language interface — let the user speak plainly
- Real execution, not simulation — perform actual device actions
- Verification matters — check results when possible
- Practical over flashy — useful automation beats gimmicks
GhostHand is powerful, but not magic.
Current limitations include:
- some actions depend on device/vendor behavior
- some apps do not expose stable deep links or intent APIs
- external messaging actions require careful human confirmation
- Korean text input via ADB can be unreliable in some contexts
- wireless ADB sessions may need reconnection
- some Android build tools need special Termux-specific workarounds
One major GhostHand milestone was proving that Android APKs can be built directly in Termux on-device.
A key workaround was forcing Gradle to use the Termux-installed aapt2 binary instead of the default Linux AAPT2 binary downloaded by Android Gradle Plugin.
This enabled successful APK builds for simple Android apps entirely on the phone.
GhostHand is an exploration of what happens when a smartphone gains a local AI "ghost hand" that can:
- understand intent
- act on the device
- verify outcomes
- build tools for itself
- and gradually become a practical personal operator layer
GhostHand는 안드로이드 폰 안의 AI 에이전트가, 같은 폰을 다시 제어하는 실험적 프로젝트입니다.
핵심 구성은 다음과 같습니다:
- Termux: 로컬 실행 환경
- OpenClaw: AI 에이전트 런타임
- 무선 ADB: 안드로이드 제어 브리지
- 자연어 명령: 사용자 인터페이스
즉, 사용자가 채팅으로 명령하면 GhostHand가 이를 해석해서 앱 실행, 설정 열기, 손전등 제어, APK 빌드/설치 같은 실제 동작으로 연결하는 구조입니다.
MIT