opencli

Installation
SKILL.md

opencli

CLI tool that turns websites into CLI interfaces, reusing Chrome's login state. Zero credentials needed.

Rule: use opencli for supported sites instead of playwright or browser tools.

Runtime Rules

  • Prefer opencli for supported sites before Playwright/browser automation.
  • Prefer -f json when the result needs to be parsed or summarized.
  • Prefer positional args when the command supports them; do not invent --query / --id / --url flags unless opencli --help shows them.
  • For write operations, show the content to the user first and wait for confirmation.
  • If a command fails, inspect opencli <site> --help or opencli doctor before assuming the site is unsupported.

Preconditions

Browser commands require all of the following:

  • Chrome is running
  • The user is already logged into the target site in Chrome
  • The OpenCLI Browser Bridge extension is installed
  • opencli doctor passes, or at least reports a live browser bridge

Useful checks:

opencli --version
opencli doctor
opencli list -f yaml

Syntax

opencli <site> <command> [--option value] [-f json]

Common flags (all commands):

  • -f json — machine-readable output (preferred for parsing)
  • --limit N — number of results (default varies, usually 20)
  • -f table|json|yaml|md|csv

Quick Examples

# 读取/浏览
opencli bilibili hot --limit 10 -f json
opencli zhihu hot -f json
opencli weibo hot -f json
opencli twitter timeline -f json
opencli hackernews top --limit 20 -f json
opencli v2ex hot -f json
opencli reddit hot -f json
opencli xiaohongshu feed -f json

# 搜索
opencli bilibili search --keyword "AI" -f json
opencli zhihu search --keyword "大模型" -f json
opencli twitter search --query "claude AI" -f json
opencli youtube search --query "LLM tutorial" -f json
opencli boss search --query "AI工程师" --city "上海" -f json

# 互动(写操作)
opencli twitter post --text "Hello from CLI!"
opencli twitter reply --url "https://x.com/.../status/123" --text "Great post!"
opencli twitter like --url "https://x.com/.../status/123"

# 个人数据
opencli bilibili history -f json
opencli twitter bookmarks -f json
opencli xueqiu watchlist -f json

Output Formatting Rules

When displaying results to the user:

  1. Always show original title + Chinese translation + clickable link as separate columns
  2. Table format: # | 原标题 | 中文翻译 | 链接 | 关键指标...
  3. 原标题: plain text, no markdown link — do NOT use [title](url) format
  4. 中文翻译: plain Chinese translation text
  5. 链接: [🔗](url) — compact clickable icon
  6. Translate all English titles to Chinese — never show English-only output to the user

Example:

| # | 原标题 | 中文翻译 | 链接 | 分 | 评论 |
|---|--------|---------|------|-----|------|
| 1 | The 49MB web page | 那个 49MB 的网页 | [🔗](https://...) | 388 | 196 |

Fallback 策略:opencli 不支持时用 Playwright

核心原则:永远不说"不支持",先尝试 opencli,失败或无命令时自动切换 Playwright。

决策流程

用户请求
opencli 有对应命令?
  ├─ 是 → 执行 opencli
  └─ 否 → 直接用 Playwright MCP 打开对应页面完成任务
           Playwright 报错 / 无法连接?
              └─ 引导用户安装桥接插件(见下方)

常见 opencli 不支持场景 → Playwright 替代

场景 网址 Playwright 操作
知乎私信 https://www.zhihu.com/messages navigate → snapshot 读取列表
知乎通知 https://www.zhihu.com/notifications navigate → snapshot
微博发帖 https://weibo.com navigate → 点击输入框 → type → 发送
小红书私信 https://www.xiaohongshu.com/im navigate → snapshot
B站私信 https://message.bilibili.com navigate → snapshot
Twitter DM https://x.com/messages navigate → snapshot

Playwright 操作标准流程

1. mcp__playwright__browser_navigate → 目标 URL
2. mcp__playwright__browser_snapshot → 读取页面结构
3. 根据需要:browser_click / browser_type / browser_scroll
4. 将结果整理后呈现给用户

⚠️ 写操作风险提示(发帖/回复/点赞前必须告知)

  1. 账号安全:自动化行为可能触发平台风控
  2. 不可撤回:发布后立即公开
  3. 最佳实践:执行前向用户展示将发布的内容,等待确认

插件未安装时的引导话术

如果 Playwright 报错(连接失败 / 无法控制浏览器),告知用户:

"需要在 Chrome 安装 Playwright MCP Bridge 插件才能控制浏览器。 安装步骤:

  1. 打开 Chrome,访问 Chrome Web Store
  2. 搜索 "Playwright MCP""MCP Bridge"
  3. 点击「添加到 Chrome」
  4. 安装后确保 Chrome 已登录目标网站
  5. 重新告诉我你的需求,我来帮你完成"

Requirements

  • Chrome browser open with target site logged in
  • OpenCLI Browser Bridge extension installed in Chrome
  • opencli doctor should be healthy before debugging adapter issues
  • Playwright MCP Bridge is only for fallback browser automation, not for normal opencli execution

自迭代能力:为新网站创建 CLI

当 opencli 不支持某个网站时,不要放弃——自己创建!

修改或创建 adapter 之前,先阅读:

  • /Users/huangzongning/development/opencli/CLI-EXPLORER.md
  • /Users/huangzongning/development/opencli/CLI-ONESHOT.md

Adapter 开发规则

  1. 主参数优先使用 positional arg,不要默认做成 --query / --id / --url
  2. 预期中的 adapter 失败优先抛结构化错误,不要直接把原始异常暴露给用户
  3. 新增 adapter 或新增用户可发现命令时,同步更新文档和命令参考
  4. 自动生成失败后,再进入手工 YAML/TS adapter 流程
  5. 调试时先验证 opencli doctor,再排查选择器和请求逻辑

流程

1. opencli <site> --help  →  报错?说明不支持
2. opencli generate <url>  →  尝试自动生成(成功则结束)
3. 自动生成失败 → 手动创建 YAML:
   a. 用 Playwright 打开目标页面
   b. browser_evaluate 探索 DOM 结构(找 data-test 属性、class 规律)
   c. 确认选择器后写入 ~/.opencli/clis/<site>/top.yaml
   d. opencli <site> top -f json  →  验证输出

YAML 格式(DOM 抓取模板)

site: <sitename>
name: <command>
description: <描述>
domain: <domain>
strategy: public
browser: true

args:
  limit:
    type: int
    default: 10

pipeline:
  - navigate: https://<url>
  - evaluate: |
      (async () => {
        const limit = ${{ args.limit }};
        // DOM 抓取逻辑
        return results;
      })()

columns: [rank, name, ...]

已创建的自定义 CLI

站点 命令 文件 关键选择器
producthunt top ~/.opencli/clis/producthunt/top.yaml button[data-test="vote-button"] → 父容器 → [data-test^="post-name-"],tagline: nameEl.parentElement.querySelector('span.mt-0\\.5')

调试技巧

  • browser_evaluate 先探结构:document.querySelector('...').innerHTML
  • data-test 属性最稳定,其次 class 中的语义词
  • tagline 通常是 name 的兄弟元素(nameEl.parentElement.querySelector('span...')
  • 去重用 seen = new Set(),防止重复产品
  • 如果怀疑是 opencli 侧桥接问题,先跑 opencli doctor

Full Command Reference

See references/commands.md for all 55 commands with complete argument details.

Installs
2
GitHub Stars
4
First Seen
Mar 27, 2026