Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add nexus-research-lab/nexus --skill computer-usegit clone --depth 1 https://github.com/nexus-research-lab/nexusWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/nexus-research-lab/nexus/computer-use)<a href="https://agentmods.dev/skills/nexus-research-lab/nexus/computer-use"><img src="https://agentmods.dev/badge/skills/nexus-research-lab/nexus/computer-use/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/nexus-research-lab/nexus/computer-use"><img src="https://agentmods.dev/badge/skills/nexus-research-lab/nexus/computer-use.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00051 | $0.01709 |
| Opus 5.5 | $0.00020 | $0.00684 |
| Sonnet 5.5 | $0.00010 | $0.00342 |
| Haiku 4.5 | $0.00005 | $0.00171 |
Grade A, and why
computer-use scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 57 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Computer Use
当用户要求你查看或操作当前电脑上的原生桌面应用时使用本 Skill。浏览器网页优先使用 Browser 能力;普通文件读写使用文件或文档能力。不要为了绕过能力边界调用 shell、文件系统、浏览器调试协议或外部 MCP。
入口
Computer Use 只通过当前回合的 nexus.command 使用。先读取固定能力与权限:
{"domain":"computer","action":"inspect"}
如果返回错误,说明能力未启用、宿主权限不足或 sidecar 不可用;把这个事实告诉用户,不要猜 endpoint、token、CLI 参数,也不要尝试重放动作。
标准流程
-
需要新的操作时先读取精确 contract:
{"domain":"computer","action":"contract","operation":"<operation>"} -
已知目标时使用
list_applications并传query(如"Chrome"或"com.google.Chrome"),避免对整个应用目录分页。query 只筛选名称/ID,不选择目标或授予权限;再用select_target以返回的精确应用名称或稳定 application id 选择目标,明确请求所需 allowed_actions、allow_foreground_input 和 ttl_seconds,由宿主确认后生效。参数上限读取 get_computer.action_limits/action_schema。所有 invoke 都需要顶层
request_id,包括只读操作。例如{"domain":"computer","action":"invoke","operation":"list_applications","request_id":"list-apps-1","input":{"query":"Chrome"}}。operation只放顶层,input只包含该操作 contract 声明的参数,不再嵌套 operation。 -
首次选择多窗口应用时先调用
list_windows,再把返回的window_ref或精确标题传给observe;单窗口应用可直接observe。后续省略window_selector会沿用仍存在的已选窗口;原窗口消失且有多个新窗口时再选择。window_selector是工具参数,不是要求用户打开窗口选择弹窗。 -
每次动作必须使用最新观察;首次调用
observe,后续可使用动作返回的新observation。截图或无障碍树过期后重新观察。需要同应用菜单与弹窗时,在能力支持下使用application_scene=true。 观察默认包含截图;点击、移动、拖拽和滚动必须使用带截图的观察,目标点均使用截图左上角为原点的0..1000归一化坐标;不要传图片像素。滚动必须给出目标 x、y,正向 delta_y 向下滚动。快捷键优先使用mod(macOS Command,Windows Control),回车使用enter。 -
动作使用
perform_action,把业务输入放在 contract 允许的 closed object 中;顶层生成新的request_id。如果结果未知,复用同一个 request_id 重试以完成对账,不能换 ID 重放。 -
动作返回
observation时直接读取它;未返回时重新observe。observation_error只代表后续读取失败,不能据此重做动作。缓存结果标记fresh_observation_required时重新观察。需要确认业务状态时调用verify_state。 普通observe的observation_id和elements在结果顶层;perform_action的新观察在observation内。截图是单独的图片内容,不要把它当作文本分页读取。 -
任务结束调用
close_target,丢弃本轮的 session、observation、element 和 pending mutation 状态。
安全边界
- 只使用宿主返回的 opaque refs;不能提交 PID、窗口句柄、AX/UIA 对象、路径、socket、token 或任意 native handle。
- 输入、文本和截图遵守 contract 返回的长度与数量上限。不要把敏感文本写入普通回复或日志。
- 常规动作使用 select_target 已批准的会话范围;退出应用及宿主标记的敏感操作仍需单独审批。被拒绝或权限不足时停止并说明原因。审批后 stale 必须重新观察,不自动重放。
- 只有明确返回
state=minimized才使用restore_window;window_status来自观察前的窗口发现,unknown/offscreen、小尺寸或台前调度缩略图都不能证明最小化。台前调度恢复使用focus_window。最小化窗口先用include_screenshot=false, accessibility="disabled"观察,再执行restore_window,恢复后截图。 session_active=true仅表示应用授权仍在,不保证已有可用于动作的 observation。观察失败或尚未观察时,先修复观察,不要反复select_target。stale_observation不是应用授权到期。失败的动作尝试也可能消耗旧观察;重新观察后根据当前画面决定下一步,不能机械重放。只有not_dispatched才能确定该尝试未执行。若连续输入再次让新观察失效,等待用户空闲再继续,不循环 observe/action,也不要求重新授权。- macOS 后台窗口(包括台前调度缩略窗口)若返回
target_unavailable且recovery_action="focus_window",对同一window_ref先做一次include_screenshot=false, accessibility="disabled"的元数据观察,再在已批准的范围内执行一次focus_window,然后重新截图和读取无障碍树。禁止用未成功的观察直接执行动作;一次恢复仍失败就报告具体错误,不能反复聚焦或重新授权。此流程不解除user_active。 - 应用未运行时先
list_launchable_applications并传相同query,再用返回的精确 launch_ref 调用launch_application;启动许可不包含后续输入权限。若应用在发现后已被用户启动,按discover_applications指引重新查询运行应用并select_target,不要重复 launch。 - macOS 两次动作之间的人工输入会等待连续空闲 2 秒(单次最多等 5 秒),保留应用授权,但输入之前的观察失效。重新观察后可继续;
busy且recovery_action="wait_for_idle_then_observe"时最多重新观察并尝试一次,持续忙碌就说明正在等用户空闲,不能无限重试。 user_active表示原生动作执行期间用户接管:停止自动动作并说明,不因等待或重新观察而恢复,不自动换 session。用户明确要求继续后再走新的 select_target 授权;可能已执行一部分的动作先对账。- session 过期重新 select_target;回合授权上限到达后开始新回合,不静默续期。
indeterminate只用原 request_id 对账一次;仍未知就停止并报告,不通过新 ID、重开 target 或 launch_application 绕过待对账状态。- 每个 mutation 只对应一个稳定 request_id;目标、动作或输入改变时必须换 ID。
- 不要自动点击系统权限弹窗。Screen Capture、Accessibility 或 Input Control 未授予时,说明用户需要在系统设置中授权后再试。
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 57 lines · 51 tokens per session scan A d203210a0a17
computer-use is a skill published in the GitHub repository nexus-research-lab/nexus (160 stars, last pushed 2d ago), licensed Apache-2.0. It adds 51 tokens to every session and 1,709 once invoked, about $0.0002 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-11-05.
computer-use compared with skills from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| computer-usenexus-research-lab/nexus | Skill | 160 | 2d ago | Apache-2.0 | A | 51 | Claude Code, Codex |
| agent-initializationPrism-Shadow/penguin-harness | Skill | 3.2k | yesterday | Apache-2.0 | A | 79 | Claude Code, Codex |
| cronshibing624/agentica | Skill | 360 | 2d ago | Apache-2.0 | A | 52 | Claude Code, Codex |
| ceo-setupsuyoumo/ClawProBench | Skill | 829 | 2mo ago | Apache-2.0 | A | 60 | Claude Code, Codex |
| commitment-triagesuyoumo/ClawProBench | Skill | 829 | 2mo ago | Apache-2.0 | A | 29 | Claude Code, Codex |
| routine-advisorsuyoumo/ClawProBench | Skill | 829 | 2mo ago | Apache-2.0 | A | 19 | Claude Code, Codex |
| delegationsuyoumo/ClawProBench | Skill | 829 | 2mo ago | Apache-2.0 | A | 25 | Claude Code, Codex |