Ghosthand is a high-performance local HTTP runtime that enables AI agents to observe and control Android devices through structured UI inspection and semantic interaction.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install ghosthand-skill
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install ghosthand-skill using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Ghosthand?
Ghosthand serves as a specialized Android agent substrate, exposing a local loopback HTTP API (typically at http://127.0.0.1:5583) for granular device control. Unlike generic automation tools, it provides a route-based interface specifically designed for AI agents using Openclaw Skills. It allows for sophisticated interaction with the Android OS, ranging from simple pings and health checks to complex UI tree analysis and coordinate-based gestures.
By leveraging Ghosthand, developers can build workflows that move beyond fragile coordinate tapping. The runtime supports multiple observation sources—including accessibility services, OCR, and hybrid modes—ensuring that agents can maintain situational awareness even when standard UI hierarchies are obscured or unavailable.
Ghosthand Use Cases
- Automating repetitive tasks on Android devices via AI-driven Openclaw Skills.
- Performing structured UI testing and inspection through a machine-readable API.
- Enabling semantic interaction with mobile applications using text, description, or resource ID selectors.
- Debugging Android application state and accessibility node hierarchies in real-time.
- Bridging the gap between LLM reasoning and physical device execution for mobile automation.
How Ghosthand Works
- Initialization: The agent first verifies connectivity by calling
/pingand checks system health and capability authorization via/state. - Surface Observation: The agent requests the current UI state using
/screen, defaulting to accessibility-based inspection to identify actionable elements. - Target Selection: Using the returned data, the agent selects the best interaction strategy (e.g., text-based click, ID-based find, or coordinate tap).
- Execution: The agent sends a command to the appropriate Ghosthand route, such as
/click,/input, or/scroll, to perform the action. - Verification: After an action, the agent calls
/waitor re-inspects the screen to confirm the UI has transitioned to the desired state.
Ghosthand Setup
Ghosthand runs as a local HTTP server on the target Android device. Ensure the device is reachable via loopback and the service is active.
# Check if Ghosthand is alive
curl http://127.0.0.1:5583/ping
# Verify capability state (Accessibility, Screenshot, etc.)
curl http://127.0.0.1:5583/state
# List all available commands and routes
curl http://127.0.0.1:5583/commands
Ghosthand Data Schema & Taxonomy
Ghosthand organizes interaction data through specialized routes. Below is the primary metadata and route taxonomy:
| Category | Primary Routes | Description |
|---|---|---|
| Runtime | /state, /capabilities |
Health, permissions, and available features. |
| Inspection | /screen, /tree, /find |
Structured UI data (JSON) and hierarchy. |
| Action | /click, /tap, /input, /scroll |
Execution of physical or semantic interactions. |
| Sensing | /screenshot, /clipboard |
Visual and data transport primitives. |
Note: Node IDs provided in responses are snapshot-scoped and ephemeral; they should be re-validated after any UI change.
Ghosthand Advanced Features
- Multi-Source Vision: Fallback from accessibility to
/screen?source=ocrorsource=hybridwhen UI elements are unmapped. - Semantic Resolution: Automatic clickable ancestor resolution via
/clickto handle wrapper elements efficiently. - State Settle Logic: Intelligent
/waitfunctionality that monitors for UI stability before proceeding with the next Openclaw Skills step. - Policy-Gated Capabilities: Granular control over which system permissions (e.g., notification access, screenshot) are exposed to the agent.
- IME Confirmation: Support for on-screen keyboard confirmation actions to handle input submission where standard 'Enter' keys fail.
SKILL.md
Loading
Related Openclaw Skills
RizzForms Form Backend Skill
A serverless form backend manager that enables automated form creation, submission handling, and webhook delivery via a dedicated CLI.
Rapid OCR Reader
A high-performance OCR tool powered by RapidOCR to extract structured Chinese and English text from any image file.
Revenue Validator
A high-speed validation engine for AI agents to pressure-test side hustle ideas against real-world market demand.
Context Manager Skill
A sophisticated long-term memory management tool for AI agents that utilizes dual-layer vector and timeline indexing to optimize context retrieval.
Gaokao English Vocabulary Frequency Grading System
An interactive, data-driven web application for mastering Gaokao English vocabulary through frequency-based classification and dark-themed UI.
Gold Monitor CN
A comprehensive monitoring tool for the Chinese gold ecosystem, tracking ETFs, spot prices, international futures, and mining stocks.
Featured*
