被低估的 Agent 必装配套:OpenCLIUnderrated Agent Accessory: OpenCLI


AI free token/AI免费积分

被低估的 Agent 必装配套:OpenCLI

被低估的 Agent 必装配套:OpenCLI

你的 AI Agent 能上网搜资料,但你有没有想过——它访问的每个网站,都得从头摸索一遍?

知乎的搜索框叫什么名字?小红书的翻页逻辑是什么?每次换一个网站,AI 就要重新研究一遍。OpenCLI 做了一件事:把 148 个网站的操作预先封装好,AI 只需要发一条命令。而且 148 不是上限——没封装的网站,AI 可以自己现写。 这才是其他浏览器自动化方案真正缺的东西。

过去几周,我把 OpenCLI 装进了日常的 AI 工作流,从搜索资料到抓取文章、从多平台情报收集到小红书内容采集。结论是:它让 AI 获取网络信息这件事,从“每次都重新摸索”变成了“一条命令就走”。


一、你的 AI 不缺浏览器,缺的是“知道怎么用每个网站”

现在的 AI Agent 大多配了上网能力——Claude Code 能搜网页,Codex 能开浏览器,各类 Agent 框架也都接入了搜索引擎。但你有没发现,真正让它去特定网站上找东西的时候,体验很差?

比如你让 AI “帮我把这篇微信公众号文章的内容整理出来”。AI 会怎么做?

它没有“微信文章一键下载”这个能力。它只能打开一个空白浏览器,自己研究怎么访问这篇文章、怎么提取正文、怎么下载图片。这个过程通常有两种结果:要么折腾半天拿到的内容残缺不全,要么直接被微信的安全系统挡在门外。

问题不在于 AI Agent “能不能上网”——几乎所有 AI Agent 都能上网。问题在于,它们对每个网站都是“第一次来”。 就像一个游客到了一个陌生城市,每去一个地方都得重新查地图、研究交通、摸索规则。

这就是大多数浏览器自动化方案的根本局限:它们给你的是一个空浏览器。目标是“让 AI 能操作浏览器”,但交付的是“让 AI 自己研究怎么操作每个网站”。

OpenCLI 走了另一条路。


二、148 个网站,一条命令——这才是“稳定”的来源

OpenCLI 的核心思路不是“给 AI 一个浏览器”,而是“给每个网站一个遥控器”。

你在 Chrome 上装一个扩展(就跟装广告拦截插件一样简单),然后装一个命令工具。之后,你的 AI 访问任何已支持的网站,只需要发一条命令:

# 下载微信公众号文章(完整保留正文和图片)
opencli weixin download --url "文章链接" --output ./articles

# 搜索小红书
opencli xiaohongshu search "露营装备" --limit 10

# 查看 Hacker News 热门
opencli hackernews top --limit 5

# 搜索知乎
opencli zhihu search "大模型部署"

# 搜索 Reddit
opencli reddit search "AI agent" --subreddit MachineLearning

# 搜索 arXiv 论文
opencli arxiv search "attention mechanism"

# 抓取 36kr 快讯
opencli 36kr search "AI agent"

关键不是命令本身,而是每条命令背后都有一个预写好的“适配器”。

适配器是什么?就是社区志愿者针对每个网站写好的一份操作指南。它里面包含了这个网站的登录方式、搜索框在哪里、翻页逻辑怎么处理、正文怎么提取、图片怎么下载。AI 不需要知道这些——它只需要执行命令,然后收结果。

这就是 OpenCLI 和其他方案最本质的区别:其他方案让 AI 自己去理解每个网站,OpenCLI 让社区替 AI 先做了这件事——但没做完的,你也可以自己补。

对 AI 来说,148 个网站变成了 148 个按钮。想看知乎?按一下。想看微信公众号?按一下。结果是稳定、可预测的——每次返回的格式都一样,不会因为 AI 今天“理解错了网页结构”而拿到一堆乱码。

而且 OpenCLI 的名字本身就说清楚了它的设计思路:把任何网站、任何软件都变成一条 CLI 命令。 除了 148 个预置的网站适配器,它还可以桥接你电脑上已有的命令行工具(比如 opencli gh 直接调 GitHub CLI、opencli docker 管理容器),甚至操作桌面软件(Cursor、Codex 这类 Electron 应用也能用 opencli cursor 来驱动)。对 AI 来说,这意味着它只需要学会一种交互模式——发命令、收结果——就能操控浏览器、终端工具、桌面应用,而不需要为每种东西学一套新操作。

你可能会问两个问题:万一网站改版了适配器失效怎么办?万一我要用的网站没有现成的适配器怎么办?

第一个问题:有 148 个适配器,哪怕 10% 在某天坏了,剩下 130 多个仍然正常工作。而且修一个已有的比从零造一个简单得多——工具内置了自动诊断,能告诉你是哪里出了问题。

第二个问题更重要:没有适配器,不代表用不了。 OpenCLI 本身就支持 Agent 自己写适配器——它的 opencli browser 命令可以让 AI 直接操作任意网页,然后把自己摸索出来的操作逻辑保存成一个新的适配器,下次直接用。换句话说,148 不是上限,是起点。你需要的网站没适配器?让 AI 当场给你写一个。


三、关键在哪儿:用的就是你自己的浏览器

适配器解决了“怎么操作”,但还有一个更根本的问题:谁来操作?

大多数浏览器自动化工具的做法是:在后台开一个全新的、你看不见的浏览器窗口,让 AI 在里面干活。这带来两个麻烦:第一,你在这个新窗口里没有登录任何网站,每用一个新平台就得单独配一次登录信息;第二,你的数据经过了什么路径、存在哪里,不那么透明。

OpenCLI 不一样。它不创建新浏览器,而是直接在你正在用的 Chrome 里干活——就是你每天刷知乎、看小红书、登录微信的那个 Chrome。你装一个扩展,它就多了一个“听从 AI 指令”的能力,仅此而已。

这带来三个好处:

  • 不用反复登录。 你在 Chrome 里登过的网站,AI 直接用。不需要为每个网站单独配置账号密码。
  • 获取数据不经过第三方中转服务。 网页内容的抓取在本地浏览器里直接完成,不像有些方案那样先经过别人的服务器再转发给你。
  • 更难被网页前端识别为自动化。 普通浏览器自动化工具在网页看来有明显的“机器人特征”——浏览器会主动告诉网页“我是被程序控制的”,显卡指纹是空的,加密握手特征也和正常浏览器不同。OpenCLI 走的是你真实 Chrome,这些标记全都不存在。当然,这不等于能绕过网站的服务器端风控(频率和行为的监控),但在“网页看到的信息”这一层,确实更干净。

四、让 Agent 替你刷信息流:省时间,也省"算法绑架"

工具本身聊完了,下面说说它能干什么。

每天打开手机刷知乎热榜、看 Hacker News 头条、扫一眼 Reddit 上某个板块的讨论、翻翻 36kr 的 AI 快讯……这些动作占了多少时间?而且你看到的内容,是平台算法帮你挑的——它知道你会点什么、会停在哪条上,然后喂更多类似的东西给你。

OpenCLI 可以反过来:让 Agent 按你的规则去各平台拉信息,而不是按算法的规则。

每天早上一杯咖啡的时间,Agent 已经在后台跑完了——HN 前十的热帖标题和链接、Reddit 某个子版块的新讨论、知乎某个话题下的最新回答、36kr 的 AI 资讯。全都是原始数据,没有算法排序,没有“猜你喜欢”。

你要的不是一个更聪明的推荐引擎,而是一个不受算法影响的信息直通车。


五、不止搜索:微信公众号、知乎、Reddit……全是一个套路

上面聊了日常浏览,再说说 OpenCLI 更主动的用法——按需搜索。

举个最常见的场景:你想让 AI 帮你收集某个技术话题在各平台上的讨论。以前的做法是什么?自己一个个打开网页、复制粘贴、整理格式。有了 OpenCLI 之后:

# 技术话题多平台调研
opencli hackernews search "deepseek" --limit 10          # HN 讨论
opencli reddit search "deepseek v4" --subreddit LocalLLaMA  # Reddit
opencli zhihu search "DeepSeek 部署"                     # 知乎
opencli 36kr search "DeepSeek"                           # 36kr 快讯
opencli arxiv search "deepseek"                          # 学术论文

五个平台,五条命令,流程完全一致。

这背后是一个更通用的模式:动笔之前,先让 Agent 跑一圈各平台做功课。 写行业报告,让它去 HN、Reddit、知乎、36kr 看看有没有相关讨论;写自媒体,让它去社交平台上翻翻用户最近在关注什么、争议什么——不只是看爆款标题,把评论区拉出来读一遍,反复出现的痛点和疑问才是真实用户在意的信号。比坐在电脑前凭空构思选题管用得多。


六、给你的 AI 装上它

OpenCLI 不是万能的。

它需要你的电脑开着 Chrome(如果你用远程服务器,公开网站没问题,需要登录的那种就用不了)。它的操作速度不算快——每个命令跑在真实浏览器里,通常需要十几二十秒。148 个适配器是社区维护的,网站改版了可能会失效(但修起来不复杂)。

但它解决了一个其他方案没认真对待的问题:让你的 AI 每次访问同一个网站时,得到的是稳定、可预测、格式一致的结果。 不用每次重新摸索,不用祈祷今天网页结构没变,不用手动整理格式。

安装很简单:把 https://github.com/jackwener/opencli[1] 丢给你的 Agent,说“帮我装上”。它会自己搞定安装、启动、检查连通性。你只需要在 Chrome 应用商店装一个同名扩展,就完事了。

一个简单的自检清单:

  • 你的 AI 能稳定地获取微信公众号文章吗?
  • 你的 AI 能在知乎、Reddit、Hacker News 上搜索内容吗?
  • 你的 AI 访问不同网站时,是每次都重新摸索,还是有固定的“套路”?
  • 你的 AI 获取网络信息时,数据经过了第三方服务器吗?

如果任何一个问题让你犹豫——OpenCLI 值得花十分钟试一下。


本文提到的项目:OpenCLI(github.com/jackwener/opencli)。

 

本文由AI辅助创作,作者对内容负责

引用链接

[1]https://github.com/jackwener/opencli

 

Underrated Agent Accessory: OpenCLI

The underrated must-have companion for Agents: OpenCLI

Your AI Agent can search the web, but have you ever thought——every website it visits, it has to figure out from scratch?

What's the name of Zhihu's search box? What's Xiaohongshu's pagination logic? Every time it switches websites, the AI has to research it all over again. OpenCLI did one thing: pre-packaged operations for 148 websites, so the AI only needs to send one command. And 148 isn't the limit—for websites not packaged, the AI can write them on the fly. This is what other browser automation solutions truly lack.

Over the past few weeks, I've integrated OpenCLI into my daily AI workflow, from searching for information to scraping articles, from multi-platform intelligence gathering to Xiaohongshu content collection. The conclusion: it turned AI's access to online information from "figuring things out from scratch every time" to "one command and it's done."

1. Your AI doesn't lack a browser; it lacks "knowing how to use each website"

Most AI Agents today come with web access—Claude Code can search web pages, Codex can open a browser, and various Agent frameworks have integrated search engines. But have you noticed that when you actually ask it to find something on a specific website, the experience is terrible?

For example, you ask the AI to "help me organize the content of this WeChat official account article." What does the AI do?

It doesn't have the ability to "one-click download WeChat articles." It can only open a blank browser and figure out how to access the article, extract the body text, and download images. This process usually ends in one of two ways: either you get incomplete content after wasting a lot of time, or it's blocked outright by WeChat's security system.

The problem isn't whether AI Agents "can access the internet"—almost all AI Agents can. The problem is that they're "first-time visitors" to every website. Like a tourist arriving in an unfamiliar city, every new place requires re-checking the map, figuring out transportation, and learning the rules.

This is the fundamental limitation of most browser automation solutions: they give you an empty browser. The goal is "let the AI operate a browser," but what's delivered is "let the AI figure out how to operate each website by itself."

OpenCLI took a different path.

2. 148 websites, one command—this is where "stability" comes from

OpenCLI's core idea isn't "give the AI a browser," but "give each website a remote control."

You install an extension on Chrome (as simple as installing an ad blocker), then install a command-line tool. After that, when your AI visits any supported website, it only needs to send one command:

The key isn't the command itself, but the pre-written "adapter" behind each command.

What is an adapter? It's an operation guide written by community volunteers for each website. It contains the site's login method, where the search box is, how pagination works, how to extract the main text, and how to download images. The AI doesn't need to know any of this—it just executes the command and receives the results.

This is the most essential difference between OpenCLI and other solutions: other solutions make the AI understand each website by itself, while OpenCLI has the community do this for the AI in advance—but what hasn't been done, you can complete yourself.

For the AI, 148 websites become 148 buttons. Want to see Zhihu? Press once. Want to see WeChat official account articles? Press once. The results are stable and predictable—the format returned is the same every time, and you won't get garbage just because the AI "misunderstood the page structure" today.

And OpenCLI's name itself makes its design philosophy clear: turn any website and any software into a CLI command. Beyond the 148 pre-built website adapters, it can also bridge existing command-line tools on your computer (e.g., opencli gh directly calls GitHub CLI, opencli docker manages containers), and even drive desktop software (Electron apps like Cursor and Codex can be driven with opencli cursor). For the AI, this means it only needs to learn one interaction pattern—send a command, receive results—to control browsers, terminal tools, and desktop applications, without having to learn a new set of operations for each one.

You might ask two questions: what if a website redesigns and the adapter stops working? What if the website I want to use doesn't have an existing adapter?

First question: with 148 adapters, even if 10% break on a given day, the remaining 130+ still work normally. And fixing an existing one is much easier than building one from scratch—the tool has built-in automatic diagnostics that can tell you where the problem is.

The second question is more important: not having an adapter doesn't mean it can't be used. OpenCLI itself supports Agents writing their own adapters—its opencli browser command lets the AI directly operate any web page, then save the operation logic it figured out as a new adapter for direct use next time. In other words, 148 isn't the ceiling; it's the starting point. The website you need has no adapter? Have the AI write one for you on the spot.

3. Where's the key: it uses your own browser

Adapters solve "how to operate," but there's an even more fundamental question: who operates?

Most browser automation tools work by opening a brand-new browser window in the background that you can't see, and letting the AI work inside it. This creates two problems: first, you're not logged into any website in that new window, so every new platform requires configuring login information separately; second, the path your data takes and where it's stored isn't very transparent.

OpenCLI is different. It doesn't create a new browser; it works directly in the Chrome you're already using—the same Chrome where you browse Zhihu, scroll Xiaohongshu, and log into WeChat every day. You install an extension, and it gains the ability to "obey AI commands," nothing more.

This brings three benefits:

No repeated logins. The websites you've logged into in Chrome, the AI uses directly. No need to configure account credentials for each website separately.

Data retrieval doesn't go through third-party relay services. Web content scraping is done directly in your local browser, unlike some solutions that route data through someone else's server before forwarding it to you.

Harder for web frontends to identify as automation. Ordinary browser automation tools show obvious "bot characteristics" to websites—the browser actively tells the page "I'm controlled by a program," the GPU fingerprint is empty, and the TLS handshake characteristics differ from a normal browser. OpenCLI uses your real Chrome, so none of these markers exist. Of course, this doesn't mean it can bypass server-side risk control (frequency and behavior monitoring), but at the "what the webpage sees" level, it is indeed cleaner.

4. Let the Agent browse your information feeds for you: save time and escape "algorithmic hijacking"

Now that we've covered the tool itself, let's talk about what it can do.

Every day you open your phone to scroll Zhihu's hot list, check Hacker News headlines, glance at discussions in a Reddit subreddit, flip through 36kr's AI news... how much time do these actions take? And the content you see is picked by the platform's algorithm—it knows what you'll click and where you'll stop, then feeds you more of the same.

OpenCLI can flip this: let the Agent pull information from various platforms according to your rules, not the algorithm's rules.

In the time it takes to have a cup of coffee in the morning, the Agent has already finished running in the background—top 10 hot post titles and links on HN, new discussions in a Reddit subreddit, latest answers under a Zhihu topic, and AI news from 36kr. All raw data, no algorithmic ranking, no "recommended for you."

What you want isn't a smarter recommendation engine, but an information express lane that isn't influenced by algorithms.

5. More than search: WeChat official accounts, Zhihu, Reddit... all the same pattern

We've talked about daily browsing above; now let's discuss OpenCLI's more proactive use—on-demand search.

Take the most common scenario: you want the AI to collect discussions on a technical topic across platforms. What was the old approach? Opening web pages one by one, copy-pasting, and formatting. With OpenCLI:

Five platforms, five commands, identical process.

Behind this is a more general pattern: before writing, have the Agent make a round of the platforms to do its homework. Writing an industry report? Have it check HN, Reddit, Zhihu, and 36kr for relevant discussions. Writing self-media content? Have it browse social platforms to see what users are recently paying attention to and debating—don't just look at viral headlines; pull up the comment sections and read through them. Repeatedly appearing pain points and questions are the real signals users care about. This is far more useful than sitting at a computer and brainstorming topics out of thin air.

6. Install it on your AI

OpenCLI isn't a silver bullet.

It requires Chrome to be running on your computer (if you use a remote server, public websites are fine, but ones requiring login won't work). Its operation speed isn't fast—each command runs in a real browser and usually takes ten to twenty seconds. The 148 adapters are community-maintained, so they may break when websites are redesigned (but fixing them isn't complicated).

But it solves a problem other solutions haven't taken seriously: when your AI visits the same website each time, it gets stable, predictable, consistently formatted results. No need to re-figure things out every time, no need to pray the page structure hasn't changed today, no need to manually format the output.

Installation is simple: throw https://github.com/jackwener/opencli [1] at your Agent and say "install it for me." It will handle the installation, launch, and connectivity check by itself. All you need to do is install an extension with the same name from the Chrome Web Store, and you're done.

A simple self-check checklist:

Can your AI reliably retrieve WeChat official account articles?

Can your AI search content on Zhihu, Reddit, and Hacker News?

When your AI visits different websites, does it re-figure things out every time, or does it have a fixed "playbook"?

When your AI retrieves online information, does the data pass through third-party servers?

If any of these questions make you hesitate—OpenCLI is worth a ten-minute try.

Project mentioned in this article: OpenCLI (github.com/jackwener/opencli).

This article was AI-assisted and the author is responsible for the content.

Reference links

[1] https://github.com/jackwener/opencli

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry