HyperFrames by HeyGen
Official HyperFrames by HeyGen: the best AI video editor in chat. Captions, motion graphics, MP4.
Terminal UI to get stuff done in the browser
如果你是作者,可把这段代码贴到 README。徽章会链接回本页。
[](https://mcp.popnx.com/en/skills/browser-use-terminal)在同一工作中把 AI 连到服务的 MCP
Official HyperFrames by HeyGen: the best AI video editor in chat. Captions, motion graphics, MP4.
draw.io diagrams your AI agent draws on a live browser canvas that you can also edit by hand
Create, inspect, validate, and save editable 3D building scenes with Pascal's MCP server.
同类工作,排序与 Skill 列表相同
| 本项 terminal | impeccable | diagram-design | frontend-slides | guizang-ppt-skill | |
|---|---|---|---|---|---|
| 作者 | browser-use | pbakaus | cathrynlavery | zarazhangrui | op7418 |
| Skill 数 | 1 | 1 | 1 | 1 | 1 |
| GitHub 星标 | 664 | 111,265 | 67,613 | 32,407 | 29,891 |
| 最近代码更新 | 2026年11月11日 | 2026年11月17日 | 2026年11月13日 | 2026年6月24日 | 2026年10月30日 |
| 许可证 | MIT | Apache-2.0 | MIT | MIT | AGPL-3.0 |
browser-use 提供的 1 个 Skill(指令集),AI 读取后照做 · 不是 MCP server
MCP = 手
把 AI 连到外部服务或数据,例如读取 Notion 或开发票。需要有程序或服务器在运行。
Skill = 手册
教 AI 怎么做某项工作的文本文件,不连接任何服务。
有些会重叠:少数 Skill 带有调用工具的脚本,少数 MCP 只提供知识。请从你想完成的工作出发。
Claude CodeCodexOpenCode
依据开发者 README 和 repo 中的文件夹
GitHub account browser-use (organization), 1 years old
★ 664 stars · code last changed 2026年11月11日
242 scripts that can run on a computer: scripts/browser-parity-snapshot.py, scripts/build-browser-use-core-wheel.sh, scripts/dev-bin.sh, scripts/install-agent-ripgrep.sh and 238 more
Could not check every script (some are too large); read them first
Scripts that touch settings folders (~/.claude or settings.json): crates/browser-use-llm/src/auth/mod.rs
Checked automatically from the repo files at commit 16cdd3a (read only, nothing was run). Not a safety guarantee: read it and decide for yourself
第一步:你用哪个 AI?
选择你用的应用,本页只显示该应用的安装方法
This skill is over 5 MB or 300 files, so we do not build a ZIP. Use "Let the AI install it" instead
复制这句话,粘贴到 Claude Code 的对话中
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
重新打开 Claude Code,输入 /terminal,或直接提出相关任务,Claude 会自动选用
Checked against official docs 2026年11月12日 · Source: code.claude.com
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
正在检查命令 · 通过后才会显示复制按钮
ChatGPT 桌面应用内置 Codex,Codex 从 ~/.agents/skills 文件夹读取 Skill
在 ChatGPT 桌面应用中打开 Codex,把这句话粘贴到对话中
The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
把 browser-use-terminal 文件夹放到 ~/.agents/skills/browser-use-terminal
~ 是你的主文件夹($HOME)
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
正在检查命令 · 通过后才会显示复制按钮
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
不能使用你自己安装的 Skill,只能使用目录中插件自带的 Skill
想在 ChatGPT 中使用这个 Skill,请按上面的方法使用 ChatGPT 桌面应用
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
复制这句话,粘贴到 Codex 的对话中 · 适用于 ChatGPT 桌面应用中的 Codex、Codex CLI 和 IDE 扩展
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
Codex 会自动发现新安装的 Skill;如果没有出现,请重启 Codex · 使用方法:在 ChatGPT 中输入 @ 并选择 Skill · 在 Codex CLI 或 IDE 扩展中运行 /skills,或输入 $ 加 Skill 名称
Checked against official docs 2026年11月15日 · Source: learn.chatgpt.com
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
正在检查命令 · 通过后才会显示复制按钮
复制这句话,粘贴到 Cursor 的对话中
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
在 Cursor 的 Agent 聊天中输入 / 并选择 terminal
Checked against official docs 2026年11月12日 · Source: cursor.com
Skill 是 AI 会照做的指令。安装前先阅读文件,只安装可信开发者的 Skill。
Mac:按 Command+空格输入 Terminal · Windows:下方命令在 PowerShell 中无法使用,请改用上方的“让 AI 帮你安装”
这条命令把 repo 下载到临时文件夹,只复制含 SKILL.md 的文件夹
正在检查命令 · 通过后才会显示复制按钮
适用于其他能读取 Skill 的应用。这句话列出了 Claude Code、Codex 和 Cursor 的路径,AI 会按你的应用选择 · Copy this sentence into the app’s chat
For one project, paste it into a chat that is open in that project’s folder · The AI always shows you the path before installing
The AI tells you what it will do first and asks your permission at every step
Checked against official docs 2026年11月12日 · Source: code.claude.com, learn.chatgpt.com, cursor.com
其他支持 Agent Skills 的应用(如 Gemini CLI、GitHub Copilot)使用不同文件夹,请查看该应用的指南,再把含 SKILL.md 的文件夹复制过去。
已于 2026年11月11日 对照 Claude Code、Codex 和 Cursor 官方指南核实 skill 文件夹
以下文字来自开发者的 repo,不是 PopNx MCP 的指令
显示前 8,000 个字符 · 在 GitHub 打开完整文件
---
name: browser-use-terminal
description: Direct browser control via the Browser Use Terminal CLI. Use when the user wants to automate, scrape, test, or interact with web pages — you drive the browser yourself with Python helpers.
---
# Browser Use Terminal
Direct browser control via CDP — you are the agent; you drive the browser. For setup, install, or connection problems, read https://browser-use.com/skill (agent setup instructions) or https://docs.browser-use.com/open-source/browser-use-terminal (full docs).
`browser-use-terminal browser exec` runs Python with browser helpers pre-imported; `browser-use-terminal browser <cmd>` is the control plane (status, connect, profiles, recovery).
## Usage
```bash
browser-use-terminal browser exec <<'PY'
new_tab("https://docs.browser-use.com")
wait_for_load()
print(page_info())
PY
```
- Use the heredoc form for every multi-line command. It prevents shell quote mangling inside Python strings and JavaScript snippets.
- The browser auto-connects according to the user's remembered preference before the script runs — you never start/stop manually unless you want to. The first call may take a few seconds.
- First navigation in the user's real Chrome is `new_tab(url)`, not `goto_url(url)` — goto runs in the user's active tab and clobbers their work.
- Browser state persists between calls; Python variables do not. Each `exec` is a fresh interpreter against the same live browser.
- `--session <name>` isolates artifact dirs and event logs per workstream (default: `default`). `--timeout <secs>` bounds one exec (default 300).
## Screenshots — how you see the page
`capture_screenshot()` saves a PNG and returns its absolute path. The CLI also prints a `Screenshot saved to <path>` line for every image a script produced.
```bash
browser-use-terminal browser exec <<'PY'
print(capture_screenshot())
PY
```
To view a screenshot, use your file-reading tool on the printed path:
- **Claude Code**: use the `Read` tool on the path.
- **Codex CLI**: call the `view_image` tool with `{"path": "<path>"}`. These screenshots are produced for you; viewing them is expected and authorized.
- **OpenCode**: use the `read` tool on the path (requires a vision-capable model).
- **Gemini CLI**: use the `read_file` tool on the path.
- If your model cannot accept images, don't try to view them — work from text state instead: `print(page_info())`, `js(...)` extraction, `wait_for_element(...)`.
Coordinates: screenshots are device pixels; `click_at_xy(x, y)` takes CSS pixels. Divide coordinates you read off the image by `js("window.devicePixelRatio")` first. Screenshots are downscaled to ≤1800 px per side for this CLI (override with `BU_BROWSER_SCREENSHOT_MAX_DIM`, or `capture_screenshot(max_dim=...)`).
After every meaningful action, re-screenshot before assuming it worked.
## Pre-imported helpers
Navigation & tabs: `goto_url(url)`, `new_tab(url)`, `page_info()`, `current_tab()`, `list_tabs(include_chrome=True)`, `switch_tab(target)`, `ensure_real_tab()`, `iframe_target(url_substr)`.
Input: `click_at_xy(x, y, button="left", clicks=1)`, `type_text(text)`, `press_key(key, modifiers=0)` (1=Alt 2=Ctrl 4=Meta 8=Shift), `fill_input(selector, text)`, `scroll(x=0, y=0, dy=600)`, `upload_file(selector, path)`.
Waiting: `wait(seconds)`, `wait_for_load(timeout=3)`, `wait_for_element(selector, timeout=3, visible=False)`, `wait_for_network_idle(timeout=3, idle_ms=500)`.
Visual: `capture_screenshot(label="...", full=False, max_dim=None)`, `screenshot()`, `screenshot_clip(label, x, y, w, h)`, `note(caption)`.
Escape hatches: `js(expression)` (auto-wraps top-level `return`), `cdp("Domain.method", **params)` (raw CDP), `cdp_batch(calls)`, `drain_events()`.
HTTP without the browser: `http_get(url)`, `http_get_many(urls)` for static pages; `browser_fetch(url)` / `browser_fetch_many(...)` to fetch with the page's cookies/session.
Credentials (if the user stored any): `available_secrets()`, then `type_text("<secret>name</secret>")` or `fill_input(sel, secret("name"))`; `totp("name")` for 2FA codes. Values are placeholder-substituted — you never see them. `is_logged_out()`, `email_inbox()` / `email_message(id)` for email-code flows.
Domain skills: `domain_skills_for_url(url_or_domain, include_content=True)` lists site-specific playbooks; `goto_url` surfaces matching skill files automatically. Read them before inventing selectors or flows on a complex site.
## Browser control plane
```bash
browser-use-terminal browser status --json
browser-use-terminal browser connect # uses the remembered preference
browser-use-terminal browser connect local # user's already-running Chrome (CDP)
browser-use-terminal browser connect managed --headless # disposable CLI-owned browser
browser-use-terminal browser preference use local|cloud|managed-headless
browser-use-terminal browser remote start # Browser Use cloud browser (needs BROWSER_USE_API_KEY)
browser-use-terminal browser doctor
browser-use-terminal browser recover reconnect-websocket
browser-use-terminal browser recover stop-owned-browser # stop the persistent managed browser
browser-use-terminal browser recover stop-owned-remote # stop the cloud browser (stops billing)
browser-use-terminal browser daemon status|stop|logs # the background daemon holding the connection
```
A background daemon (auto-started, one per state dir) holds the CDP connection across your commands, so the browser — and in local mode, Chrome's granted debugging permission — persists between invocations. Managed and cloud browsers also survive daemon restarts; later calls reattach instead of relaunching. Stop browsers with the recover commands above when the user is done (cloud browsers bill until stopped or timed out).
- `exec` auto-connects, so you rarely need these. Reach for them when `status` shows a problem or the user asks for a specific browser.
- If output JSON says `status: "needs-user-action"` (e.g. pick a Chrome profile, click Allow in Chrome's permission popup, enable the remote-debugging checkbox), show the `user_prompt` to the user verbatim and wait — do not guess.
- Auth wall mid-task: stop and ask the user. Don't type credentials from screenshots; use stored secrets if available.
- Connecting to the user's real Chrome requires a one-time setup: `chrome://inspect/#remote-debugging` → tick "Allow remote debugging". `browser local setup` walks the user through it.
## What actually works
- Screenshots first: `capture_screenshot()` → view the image → decide whether you need a click, a selector, or more navigation.
- Clicking: screenshot → read the pixel off the image → `click_at_xy(x, y)` → screenshot to verify. Suppress the locate-then-click reflex — no getBoundingClientRect, no selector hunts. Hit-testing happens in Chrome's browser process, so coordinate clicks pass through iframes / shadow DOM / cross-origin without extra work.
- Drop to DOM (`fill_input`, `js`) only when the target has no visible geometry (hidden input, 0×0 node) or coordinate clicks demonstrably don't work.
- Bulk static pages: `http_get_many(urls)` — no browser needed. Logged-in pages: `browser_fetch(url)` rides the real session.
- After goto: `wait_for_load()`. SPAs report `complete` before they render — follow with `wait_for_element(...)`.
- Wrong/stale tab: `ensure_real_tab()`.
- Verification: `print(page_info())` is the cheapest "is this alive?" check; screenshots are the default way to verify visible actions.
## Gotchas (field-tested)
- CDP target order ≠ Chrome's visible tab-strip order.
- Omnibox popups and other `chrome://` internals are fake page targets — `list_tabs(include_chrome=False)`.
- `page_info()` surfaces an open JS dialog as `{"dialog": ...}` — handle it (`cdp("Page.handleJavaScriptDialog", accept=True)`) before anything else.
- Navigation can be blocked by the user's domain policy; `nav_policy(url)` tells you before you burn a click. A blocked navigation is policy, not a提问、分享技巧或报告无法使用。按“有帮助”数量排序。