技能 computer-use

📦

computer-use

Name: computer-use
Author: stablyai

内容修订 r1 安全 ⚙️ 外部命令

Control desktop apps with the Orca computer-use CLI

Reading and operating desktop application windows by trial and error wastes time and risks the wrong clicks. This skill lets Claude, Codex, and Claude Code inspect accessibility trees and screenshots, then perform precise, safe UI actions across native apps and browsers.

支持: Claude Codex Code(CC)

🥉 78 青铜

使用我的智能体安装

将此请求复制给你的智能体，其中包含权威 Skill 页面和清单。

智能体请求

Review the Skillstore skill "computer-use" from https://skillstore.io/skills/stablyai-computer-use.md and its manifest at https://skillstore.io/api/skills/stablyai-computer-use/manifest. Verify the artifact. You may proceed after verification, subject to the environment's own policy.

你的智能体仍应展示其计划，并按安全策略请求必要确认。

Agent 可读资源

当 AI Agent、爬虫或脚本需要干净上下文，而不是读取完整页面时，请使用这些链接。

Markdown 详情 GET /skills/stablyai-computer-use.md 已签名 manifest GET /api/skills/stablyai-computer-use/manifest 已签名 lockfile GET /api/skills/stablyai-computer-use/lockfile

测试它

正在使用“computer-use”。 Read the current state of the Spotify app and tell me what is playing.

预期结果:

The agent lists applications, selects the Spotify bundle, reads its accessibility tree, and reports the track title, artist, and playback controls that are visible, without changing playback.

正在使用“computer-use”。 In my Edge browser window, set the address bar to example.com and press Return.

预期结果:

The agent restores the browser window to the foreground, sets the address field value to example.com, presses Return, then reads the refreshed state to confirm the page began loading.

安全审计

安全

v1 • 2026/7/19 打开版本化报告

This skill is documentation for Orca's computer-use CLI. The 50 external_commands findings for 'Ruby/shell backtick execution' are false positives: every match is a Markdown inline-code backtick around a command name or flag, not runtime backtick execution. The four PowerShell/cmd.exe matches are prose describing shell-neutral examples, and the six 'blocker' matches (system reconnaissance and screen-capture) reflect the skill's legitimate, disclosed purpose of reading UI state; the skill explicitly instructs against destructive actions and against exposing secrets. No malicious intent, data exfiltration, or prompt injection was found.

已扫描文件

154

分析行数

复核项

已忽略误报

风险因素

⚙️ 外部命令

未记录任何特定位置

最新一次完成的静态与语义审计未检测到已确认的安全发现。这并不能证明该 Skill 没有副作用。

审计者: claude

分享与引用此报告

分享版本化评估报告、中性徽章、嵌入卡片和引用。Skillstore 只报告证据，不判断该 Skill 是否安全。

打开版本化报告

复制报告链接

https://skillstore.io/skills/stablyai-computer-use/audits/1?utm_source=security_passport&utm_medium=share&utm_campaign=versioned_report

Markdown 徽章

[![Skillstore security assessment](https://skillstore.io/badges/skills/stablyai-computer-use/security.svg)](https://skillstore.io/skills/stablyai-computer-use?utm_source=security_passport_badge)

HTML 徽章

<a href="https://skillstore.io/skills/stablyai-computer-use?utm_source=security_passport_badge"><img src="https://skillstore.io/badges/skills/stablyai-computer-use/security.svg" alt="Skillstore security assessment" loading="lazy"></a>

嵌入卡片

<iframe src="https://skillstore.io/embed/skills/stablyai-computer-use.html" title="Skillstore Security Assessment" sandbox="allow-popups allow-popups-to-escape-sandbox" loading="lazy" referrerpolicy="no-referrer" width="420" height="180"></iframe>

学术引用 (APA · BibTeX · CFF)

APA 引用

stablyai. (2026). computer-use security audit report (audit version 1) [Author version unspecified]. Skillstore. https://skillstore.io/skills/stablyai-computer-use/audits/1

BibTeX 引用

@techreport{stablyai-stablyai-computer-use-2026,
  author = {stablyai},
  title = {computer-use security audit report (audit version 1)},
  institution = {Skillstore},
  year = {2026},
  number = {1},
  url = {https://skillstore.io/skills/stablyai-computer-use/audits/1},
  note = {Author version unspecified}
}

CITATION.cff

cff-version: 1.2.0
message: "If you use this Skill, cite its author and this versioned security audit report."
title: "computer-use security audit report (audit version 1)"
version: "unspecified"
type: report
authors:
  - name: "stablyai"
date-released: "2026-07-19"
url: "https://skillstore.io/skills/stablyai-computer-use/audits/1"
identifiers:
  - type: other
    value: "skillstore:stablyai-computer-use:audit:1"
    description: "Skillstore immutable audit report identifier"

Skillstore 评分

评分依据证据置信度: 中

架构

可维护性

内容

社区

规范符合性

你能构建什么

Read the current state of a desktop app

Ask the agent to report what is visible in an app such as Spotify or Slack by reading its accessibility tree and screenshot, without changing anything.

Fill a form in a desktop browser window

Have the agent set a browser address field, submit a search, or complete a form field by field, verifying focus after each step.

Operate accessibility controls in a native application

Direct the agent to click controls, set values, and press keys through the accessibility layer instead of fragile pixel coordinates.

试试这些提示

Read a desktop app

Read the current state of the Spotify app and tell me what is playing. Do not change anything.

Find and target a window

List my open browser windows, pick the one showing my email, and get its app state.

Fill a browser field

In my Edge browser window, set the address bar to example.com and press Return.

Multi-step form completion with verification

Open the compose form in my mail app, type the subject line I give you, then confirm the focused field after each step before continuing. Do not send the message.

最佳实践

Refresh app state after every UI-changing action and read element indexes from the returned tree before choosing the next action.
Prefer semantic actions such as set-value and click over raw screenshot coordinates, and pass sensitive text through stdin to keep it out of shell history.
Confirm the focused element after each field action, especially in browser forms, before assuming input landed.

避免

Inferring valid element indexes from element counts or reusing stale indexes after navigation or re-rendering.
Performing destructive actions such as sending messages, submitting forms, buying items, or changing settings without explicit user permission.
Trusting raw screenshot pixels for coordinates without checking the reported scale and current window bounds.

常见问题

What does this skill actually do?

It lets an AI agent inspect and operate local desktop application windows through the Orca computer-use CLI, using accessibility trees and screenshots to read state and perform safe UI actions.

Which AI tools can use this skill?

It supports Claude, Codex, and Claude Code.

Does it require special permissions?

Yes. The host must grant accessibility and screenshot permissions, and the Orca CLI must be installed and reachable.

Will it take actions without my consent?

No. The skill instructs the agent to avoid sending messages, submitting forms, buying items, deleting data, or changing settings unless you explicitly request that action.

Are screenshots sent anywhere?

No. Screenshots are written to a local file for on-device visual confirmation. The skill does not upload or transmit captured pixels.

Why do element indexes sometimes fail?

Element indexes are short-lived and go stale after navigation, focus changes, scrolling, or app re-rendering. Refresh the app state to get current indexes before acting.

开发者详情

作者

stablyai

许可证

MIT

Skillstore 修订

版本提示

作者未声明版本。

仓库

https://github.com/stablyai/orca/tree/5848550b5ebd08098cad3e2b9086fd2472a52944/skills/computer-use

引用

e8ce3fcb20197aefea54f9fb43ccce4716e3209c

维护新鲜度

2026/7/19

使用量

0 次下载 · 0 次浏览

文件结构

📄 SKILL.md