← 返回 Skills 市场

computer-use-skill

Name: computer-use-skill
Author: smallcosmos

作者 smallcosmos · GitHub ↗ · v1.0.1 · MIT-0

cross-platform ⚠ suspicious

518

总下载

当前安装

版本数

在 OpenClaw 中安装

/install computer-use-skill

功能描述

Remotely execute exact browser automation tasks by forwarding user commands directly to the Computer Use Agent (CUA) via its Python SDK.

使用说明 (SKILL.md)

Computer Use Skill

Note: This skill operates in pass-through mode - tasks are transmitted directly to CUA without modification.

Architecture

CUA Integration Path

User Task → Python SDK → CUA Sandbox → Browser Automation → Results Stream

Key Components

CUA SDK: Official Python SDK for browser automation
Virtual Environment: Isolated environment with SDK installed
Python Wrapper: Programmatic access for integration

File Structure

workspace/
├── tools/
│   └── execute_cua_task.py      # Task execution script
└── cua_venv/                    # Virtual environment with SDK

Quick Reference

Task Type	Example
Search	`"打开google页面，搜索杭州天气"`
Navigation	`"访问github.com"`
Forms	`"在登录页面输入用户名和密码"`
Screenshots	`"访问产品页面并截图"`
Scraping	`"从电商页面提取价格信息"`
Complex	`"将购物车商品结算并截图确认"`

Requirements

No external credentials handled by user
Virtual environment with CUA Python SDK installed

Usage Patterns

with run source to active cua_venv → run cd to directory tools → run python execute_cua_task.py like 'python execute_cua_task.py "在百度搜索人工智能"'

Pass-Through Protocol

Critical Rules

No task modification: Tasks are transmitted exactly as provided
No wrapping or packaging: Raw task text sent to CUA
No interpretation: CUA AI handles task understanding
No preprocessing: User text → CUA (direct path)

Workflow

User: "打开google页面，搜索杭州天气"
↓
Skill: Transmit "打开google页面，搜索杭州天气" (exact match)
↓
CUA: AI understands and executes browser automation
↓
Results: Stream of execution messages returned

What NOT to Do

Don't wrap in additional instructions
Don't process or modify user input
Don't simplify or elaborate user task

安全使用建议

Do not install or run this skill yet. Ask the publisher for: (1) the missing execute_cua_task.py and any wrapper code so you can review it, (2) the exact PyPI/package name or official URL for the 'CUA Python SDK' and verification that it's from a trusted source, and (3) documentation or a homepage for the CUA service and its sandbox endpoint. Be aware the skill's 'pass-through' rule will forward raw user text to an external agent—do not send passwords, private keys, or other secrets through prompts. If you proceed, run the code only in an isolated environment, audit the wrapper for exfiltration or hidden endpoints, and prefer a design that sanitizes or requires explicit consent before forwarding any credentials or sensitive data.

功能分析

Type: OpenClaw Skill Name: computer-use-skill Version: 1.0.1 The skill bundle defines a 'Pass-Through Protocol' in SKILL.md that explicitly instructs the AI agent to execute shell commands (python execute_cua_task.py) using raw, unmodified user input. This design pattern is highly vulnerable to shell injection attacks if the underlying script does not perform rigorous sanitization. Additionally, the core execution logic (execute_cua_task.py) and the virtual environment (cua_venv) mentioned in the documentation are missing from the provided files, preventing a full security audit of the actual execution path.

能力评估

⚠ Purpose & Capability

The SKILL.md describes forwarding user tasks to a CUA Python SDK and running tools/execute_cua_task.py inside a cua_venv, but the package contains no script, no SDK, and no homepage/source. The required files and SDK would be necessary for the stated purpose but are missing, which is inconsistent.

⚠ Instruction Scope

Instructions enforce a strict 'no modification / pass-through' policy: user text is transmitted exactly to an external CUA sandbox. That explicitly allows forwarding sensitive content (e.g., login credentials, private data) without filtering or sanitization. SKILL.md also instructs running a local Python wrapper that is not provided.

⚠ Install Mechanism

There is no install spec. SKILL.md expects a virtualenv with the 'CUA Python SDK' but gives no package name, source, or install instructions. This gap could lead an operator to pip-install an unknown package or fetch code from an untrusted source, a potential supply-chain risk.

⚠ Credentials

The skill declares 'No external credentials handled by user' yet lists actions (form fills, login) that commonly require credentials and explicitly forwards raw user input. The skill requests no env vars, but its behavior could cause users to reveal secrets in plain text tasks that will be sent to an external service—this is disproportional and contradictory.

ℹ Persistence & Privilege

always:false and no system-wide changes are requested, which is appropriate. However, the skill expects/assumes creating workspace/tools and a virtualenv—local filesystem write/execution will be needed if the agent implements the instructions.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install computer-use-skill
安装完成后，直接呼叫该 Skill 的名称或使用 /computer-use-skill 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v1.0.1

- Added .clawhub/origin.json for repository origin tracking. - Minor documentation update: usage pattern rephrased for clarity in SKILL.md. - No changes to core logic or function.

v1.0.0

- Initial release of computer-use-skill for remote browser automation via CUA. - Implements strict pass-through mode: user tasks are transmitted to CUA without any modification or preprocessing. - Integrates Python SDK and a virtual environment for isolated browser automation execution. - Provides quick reference examples for search, navigation, form input, screenshots, and web scraping. - Enforces a workflow where no user input is altered, wrapped, or interpreted—raw tasks are sent directly to CUA.

元数据

Slug computer-use-skill

版本 1.0.1

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 2

常见问题