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 agentmods add skills/metahub-tech/agent-fleet/using-visionnpx skills add metahub-tech/agent-fleet --skill using-visiongit clone --depth 1 https://github.com/metahub-tech/agent-fleetWhat 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 | $0.00077 | $0.01059 |
| Opus 5 | $0.00039 | $0.00530 |
| Sonnet 5 | $0.00015 | $0.00212 |
| Haiku 4.5 | $0.00008 | $0.00106 |
Grade A, and why
using-vision 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 — 69 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Using vision (像素级元素定位)
vision 是 pc-device(mac-device / win-device)的能力模块,补 element-action 的盲区:find_elements/tap_element 走 OS 无障碍树(AX/UIA),但网页、canvas、Electron(关 a11y)、Flutter、游戏的内容不在树里 → 它们返回空。这时用 vision 按像素定位。
何时用 vision(决策)
要点一个元素
│
├─ 原生 app 控件? → 先 find_elements / tap_element(AX/UIA,更准更省,抗布局漂移)
│
└─ 网页 / canvas / Electron / Flutter / 游戏(a11y 树拿不到)?
→ vision_locate / vision_tap(像素,本节)
不要一上来就 vision。无障碍树能拿到就用 element-action。vision 是树失效时的 fallback。
三个工具
坐标都和 core tap 同一点空间——定位完直接能点。
vision_locate(query) — 按可见文字定位
vision_locate("登录")
→ {"ok": true, "count": 2, "candidates": [
{"text": "登录", "center": [1200, 29], "box": [...], "score": 1.0, "match_field": "exact"}, ...]}
- 返回排序候选(exact > 前缀 > 包含)。先 locate 看清候选,再决定点哪个。
region=(left, top, right, bottom):只搜这块区域(强烈建议——密集页全屏 OCR ~1–2s,裁剪到亚秒)。- 没找到 →
count:0+ocr_sample(当时读到的文本),据此换词/缩 region。
vision_tap(query) — 找到即点
vision_tap("登录") # 唯一/exact 命中 → 直接点
vision_tap("hide", nth=2) # 多命中时 nth(0-based)指定第几个
vision_tap("提交", region=(300,500,460,560))
nth:0-based(0=最优候选);省略=自动(唯一或 exact 即点;多个歧义则不点、返回候选让你加 nth 或更具体的 query)。与tap_element同语义。- 歧义返回
{"ok": false, "error": "ambiguous", "candidates": [...]}→ 传 nth。
vision_locate_image(template) — 按图标图定位(无字元素)
vision_locate_image(template_b64="<截图的 base64>")
vision_locate_image(template_path="/path/on/host/icon.png", threshold=0.9)
→ {"ok": true, "found": true, "center": [x, y], "score": 0.97}
- 给一张图标/按钮的小图,返回它在屏上的中心。用于没有文字的纯图标按钮(工具栏 icon 等)。
- 单尺度:模板必须按当前显示缩放截取;跨 DPI/缩放会掉置信度(
found:false+ best_score + hint)。
红线 / 边界(重要)
- vision 只管「定位」,不是全屏 OCR、不负责「读懂页面」。 要理解页面内容、读低对比的次要文字(灰色元数据、说明文字),用
take_screenshot交给你自己的视觉——那是你的强项,vision 的 OCR 在低对比文字上会漏。 - vision 擅长高对比可交互元素(按钮/链接/标题/菜单)的精确定位(~个位 px);低对比装饰文字定位不到很正常,不是 bug。
- 模板匹配跨 DPI 是已知短板(单尺度)。
典型流程(网页点登录)
1. find_elements("登录") # web → 空(AX 树没有)
2. vision_locate("登录") # 看候选、确认 center 落在按钮上
3. vision_tap("登录") # 点中
全程 0 LLM token、离线、纯 CPU。
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 · 69 lines · 77 tokens per session scan A 71d732a6cb43
using-vision is a skill published in the GitHub repository metahub-tech/agent-fleet (2 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 77 tokens to every session and 1,059 once invoked, about $0.0004 per session on Opus 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 2026-08-31.
Other skills, from other repositories
dingtalk_channel_connect
Use a headed browser to automatically complete DingTalk channel integration for QwenPaw. Applicable when the user mentions DingTalk, developer console, Client ID, Client Secret, bot, Stream mode, binding or configuring a channel. Supports pausing when a login page is detected and resuming after the user logs in.
dingtalk_channel_connect
使用可视浏览器自动完成 QwenPaw 的钉钉频道接入。适用于用户提到钉钉、DingTalk、开发者后台、Client ID、Client Secret、机器人、Stream 模式、绑定或配置 channel 的场景;支持遇到登录页时暂停,等待用户登录后继续。.
browser_cdp
Use this skill when the user explicitly wants to connect to a running Chrome browser, scan local CDP ports, specify a cdpport, or share a single browser across multiple agents/tools. By default browser opens no debugging port; pass an explicit cdpport only when the user wants another local tool to attach.
browser_cdp
当用户明确希望连接到已运行的 Chrome 浏览器、扫描本地 CDP 端口、显式指定 cdpport,或让多个 agent / 工具共享同一个浏览器时,使用本 skill。browser 默认不开放调试端口;仅当用户明确希望其他本地工具附加时才显式传入 cdpport。.
browser_visible
当用户需要控制 browser 的浏览器启动方式时,使用本 skill。browser 默认由 Playwright 直接管理、不开放调试端口(需让其他本地工具附加时显式传 cdpport);headed 控制是否显示窗口,privatemode 保留用于兼容、不再改变默认行为,browserargs 传入额外的 Chromium 启动参数,executablepath 指定自定义浏览器可执行文件路径。.
browser_visible
Use this skill when the user needs to control the browser launch mode for browser. By default browser is managed by Playwright and opens no debugging port (pass an explicit cdpport to let another local tool attach); headed controls whether the window is visible, and privatemode is kept for backward compatibility and…