Qiaomu OpenCLI Skills: teaching an agent 79 site CLIs

Documentation for a tool this repository does not contain

Qiaomu OpenCLI Skills is a set of Claude Code skills built on top of jackwener/opencli, and the first thing to understand is that the actual engine lives elsewhere.

The project description in the README says it is a collection of Claude Code skills based on the upstream opencli project, and the badges point at that upstream repository, with a version badge for 1.6.9. The package manifest makes the relationship explicit, naming the upstream as @jackwener/opencli at version 1.6.9 and declaring it as a peer dependency on the same version.

So the architecture has two parts. The upstream package is a CLI that turns websites and desktop applications into command line interfaces, so that an AI agent can operate a browser, search content, post tweets and download videos. This repository is the layer that teaches an agent how to use that CLI, in the format Claude Code loads as skills.

The top level listing shows the shape of it, and it is documentation all the way down. There are six skill directories, qiaomu-opencli-usage, qiaomu-opencli-browser, qiaomu-opencli-explorer, qiaomu-opencli-oneshot, qiaomu-opencli-autofix and qiaomu-smart-search, plus a package.json and a LICENSE. The skill directories contain SKILL.md files, and the usage skill additionally has a commands.md with the full command reference, and the smart-search skill has a references/ directory for data source configuration.

The project is MIT licensed, the last push to the main branch was on 2026-04-09, and there are no GitHub releases. The primary language is reported as unknown, which is consistent with a repository whose content is Markdown rather than code.

Six skills, split by what the agent is doing

The skills table in the README lists the six skills with a description and a usage scenario for each, and the split is by task rather than by site.

The usage skill is the main one, described as the complete command reference for the 79 or more adapters, for everyday operations such as trending, search, posting and downloading. The browser skill covers browser automation control, for when the agent needs to drive the browser directly with clicking, typing and screenshots.

The explorer skill is a full guide to adapter creation, so it covers writing a CLI adapter for a new website. The oneshot skill is the quick version, generating a single command from a URL, so an adapter for one site without a full implementation. And the autofix skill handles broken adapters, with automatic diagnosis and repair.

The sixth, qiaomu-smart-search, is a smart search router, described as automatically choosing the best search source. Its examples show the same interface taking a query and picking a data source on its own:

# 自动路由到最佳数据源
opencli smart-search "最新 AI 论文"
opencli smart-search "React 最佳实践"

That last skill is the one that differs most in kind from the rest. The others wrap a deterministic CLI, while a router decides which adapter to call for a given question, and the core capability list describes it as choosing the best data source, with an AI source plus a dedicated source in the mix.

The autofix skill deserves a note too, because it is the one that addresses the fragility of the whole approach. Adapters are generated against the behaviour of a website, and websites change. A skill that diagnoses and repairs a broken adapter is an acknowledgement that breakage is a matter of when, not whether.

Installing the CLI, the extension and the skills

The quick start has four steps, and the second one is the part most people will not expect.

First, install the CLI globally:

npm install -g @jackwener/opencli

Second, install the browser extension. The instructions are to download opencli-extension.zip from the upstream Releases page, unzip it and load the unzipped folder at chrome://extensions, and make sure Chrome is logged in to the target site.

That third point is the design in one line. Because the CLI reuses the login state of your browser, the extension is how the CLI reaches the sites you are already signed in to, and the core capability list describes browser automation as driving Chrome directly through CDP while reusing the login state.

Third, install the skills:

# 安装所有 skills
npx skills add joeseesun/qiaomu-opencli-skills

or one at a time, by naming the skill with the --skill flag.

Fourth, verify:

opencli doctor
opencli list
opencli hackernews top --limit 5

The three verification commands are worth reading in order. doctor is the health check, list shows what is available, and the third runs a real command against a site that needs no authentication, which is a sensible smoke test because HackerNews is one of the public-API adapters.

So the dependency chain is: an npm package, a browser extension loaded unpacked, a set of skills added through npx, and the existing login state in your Chrome profile. Four things, and the third is the one that is easy to forget.

Five authentication levels, and what each one borrows

The authentication section is the most important part of the documentation for anyone deciding whether to use this, and it lists five levels.

The public level is for public APIs with no authentication. The cookie level reuses browser cookies and is described as the most commonly used. The header level needs an API token. The intercept level captures a request to obtain a token. And the ui level is CDP control of a desktop application.

The ordering is a progression from least to most invasive, and it is worth reading that way. The public level is indistinguishable from calling the site's API yourself. The cookie level borrows the session you already have in your browser, which is why the extension is a required install step. The intercept level goes further and captures a token out of a live request. The ui level is the broadest, because it drives a desktop application rather than talking to a service at all, and the core capabilities list includes desktop application control for Electron apps such as Cursor, Codex, ChatGPT, Notion and WeChat.

Two consequences follow. The first is that any of these levels above public inherits your account's permissions, so a command that posts, replies or downloads acts as you. The second is that the credentials involved are not yours in the API-key sense, so the usual rotation and revocation story does not apply in the same way.

The usage examples show the level of access the commands expect. Posting and replying exist for Twitter:

# 发推
opencli twitter post --text "Hello from CLI!"

# 回复推文
opencli twitter reply --url "https://x.com/.../status/123" --text "Great!"

And the site list is grouped by category, with social platforms including Twitter/X, Reddit, Xiaohongshu, Zhihu, Weibo, Instagram, Facebook and Bluesky, video platforms including Bilibili, YouTube, TikTok and Douyin, technical communities including HackerNews, V2EX, Stack Overflow and GitHub, AI tools including Grok, Doubao, ChatGPT, Gemini, Cursor, Codex and NotebookLM, financial sites including Yahoo Finance and Bloomberg, and a general group including Google, Wikipedia, arXiv, Medium, Substack, BOSS Zhipin and LinkedIn.

What this repository contributes, and what it does not

The contribution section is where the project states its own scope, and it is short.

It says the project is based on jackwener/opencli, that main contributions should be submitted to the upstream repository, and that this repository focuses on three things: optimising the skills documentation for Claude Code, adding Chinese usage examples, and maintaining the qiaomu prefixed versions of the skills.

That is an honest division of labour and it tells you what the fork is for. The upstream project owns the engine and the adapters. This one owns the presentation layer for an agent, in one particular agent, with examples in one particular language.

The package manifest reinforces the version coupling. The upstream is pinned at 1.6.9 with a peer dependency of ^1.6.9, the skills array lists the six skill names, and the keywords describe the same scope as the README, covering claude-code, skills, opencli, browser-automation, web-scraping, cli and the names of several individual sites.

One caution follows from that pinning. The skills document commands, and the commands come from the CLI. If the installed opencli version moves on and a command changes shape, the skill documentation drifts until someone updates it, and the drift is silent from the agent's point of view because the skill file is still what it reads. Checking the version pair before you rely on it is the cheap precaution.

The README also points to the Claude Code Skills documentation upstream, and asks readers to follow the author on X and a WeChat account for further technical writing, which is the sort of arrangement you see in the Chinese open source ecosystem where a maintainer's personal channel is part of the support model.

Editorial conclusion

Adopt Qiaomu OpenCLI Skills if you want an agent to work with real sites, because the skills encode the exact command shapes and the authentication levels so the agent does not have to guess them. Do not adopt it as the tool itself, since it holds no engine and the actual work is done by the upstream opencli package, which you install separately and which must be kept in step with the skill version. Before you rely on it, check that the skill version and the installed opencli version match, since the manifest pins the peer dependency to the same 1.6.9, and be deliberate about the authentication level you are relying on, because the cookie and intercept levels borrow a logged-in browser session rather than using an API credential of your own.

Frequently asked questions

Official sources

1. Issues
2. joeseesun/qiaomu-opencli-skills on GitHub
3. License: MIT
4. README