AI Search Hub borrows each vendor's walled garden from that vendor

The trick is asking each vendor about its own ecosystem

The design is easier to understand once you see which company stands behind each assistant. The capability table pairs every platform with its parent: Gemini with Google for search and web discovery, Grok with xAI for real-time X and Twitter search, Doubao with ByteDance for Chinese trends and content summarisation, Yuanbao with Tencent for WeChat official account search, LongCat with Meituan for Chinese knowledge and local services, Tongyi Qianwen with Alibaba for general Chinese search, and Kimi with Moonshot for long-document understanding.

Read down the parent column and the pattern is the point. ByteDance's assistant reaches Douyin and Toutiao. Tencent's reaches the WeChat ecosystem and public account articles. Alibaba's reaches Taobao. Meituan's reaches Dianping and industry reports. Each vendor has an assistant with the access its own company controls, and this Skill asks those assistants for data instead of trying to reach the same services directly.

That reframes what the project is. It is not a scraper and does not claim to be one. The stated motivation is avoiding the maintenance burden of fragile crawlers and parsing rules, one browser automation and login flow per platform, repeated logins and CAPTCHA and rate-limit and risk-control handling, and hand-stitching fragmentary results at the end. You delegate the hostile parts to whoever already fights them, and keep the aggregation.

Three lists of platforms, and an empty status column

The documentation names its platforms three separate times, and the three lists do not match.

The connected line lists eight: Gemini, Grok, Doubao, Yuanbao, LongCat, Tongyi Qianwen, MiniMax and Kimi. The support table that follows has twelve rows, adding Perplexity, Claude, Wenxin Yiyan and a final row for more vertical search sources described as an extensible surface. The capability table inside the workflow section lists seven, dropping MiniMax from the eight and adding none of the three extras.

So MiniMax appears in the connected list and the support table but not in the table that explains what each platform is good for, while Perplexity, Claude and Wenxin Yiyan appear in the support table and nowhere else. If you are planning a dependency on any of those three, the documentation does not tell you whether they work.

The deeper problem is the table's own status column. Every platform row has a 当前状态 field for current status, and every one of them is blank. There are no values, not even a marker. So the support table describes what each platform would be good at without ever stating which ones are switched on, which makes the eight-name connected line the only usable signal in the README.

The routing document is the place to resolve this, since the workflow says the agent routes a question to the most suitable platform based on its type, with details in ROUTING.md.

One question in, one merged output out

The workflow is five steps and the promise sits in the first and the last. A user or agent submits a single question, with no need to organise a separate query per platform. That question is then sent to multiple providers so each can search the data world it is best at. The platform's own native capability is reused rather than reimplemented. Results coming back from several platforms are pulled back to a single exit point, where they can be standardised, fused and consumed by a workflow. Finally the result returns to the agent or system.

The emphasis on the last step is deliberate and worth noting. The documentation is explicit that the output is not meant to stay in a browser page but becomes a reusable input inside an agent, a research system, a monitoring flow or an automation chain. That is the difference between this and a search box, and it is what makes the Skill useful in OpenClaw rather than only in a terminal.

The middle step is where the abstraction is thinnest. Standardising and fusing results from eight assistants with different output conventions is the hard part, and the README describes it in one sentence without specifying the merge format. What you consume downstream is the least documented part of the pipeline.

The README says what it is not doing, in detail

The comparison table is unusually blunt about the traditional alternative, and the admissions are the useful part. Writing your own crawler means scraping, parsing and maintenance are all yours. One platform means one automation, so every platform added is another maintenance bill. Repeatedly handling risk control and CAPTCHAs means login, rate limits and page changes keep consuming effort. Tuning keywords and retrieval strategy yourself means trial and error before results are usable. Results end up scattered and hand-assembled into a single output.

What is offered in return is described in the same five rows: reuse the platform's native search entry point, dispatch one question across multiple platforms, stay on existing entry points where possible to reduce repeated adversarial effort, borrow the platform's already-optimised ranking and understanding, and produce one unified output for an agent or workflow.

Two claims in there deserve scepticism rather than acceptance. Borrowing another company's search logic means your results inherit that company's ranking biases without your visibility into them, and staying on existing entry points means your access is exactly as stable as a consumer product's, which is not a promise a third party controls. The README also asserts free access to WeChat official account, Douyin, Weibo and Bilibili data. Free describes the Skill, not the underlying platform quotas, and nothing here should be read as those services granting third-party access on your behalf.

No install command, no dependency manifest, no licence

The repository is recorded as primarily Python, but the top level holds .gitignore, README.md, README.en.md, ROUTING.md, SKILL.md and four directories named agents/, docs/, scripts/ and no manifest at all. There is no requirements.txt, no pyproject.toml and no setup.py, so whatever the Python does, it is not installed as a package from this repository. The Skill is meant to be placed where OpenClaw reads Skills from, with SKILL.md as its entry point and ROUTING.md as the routing reference.

The visible documentation does not include an installation section, so if you are looking for a command to run, there is not one here. Read SKILL.md and ROUTING.md first, since they are the two documents the README itself points you to.

The licence situation is the thing to resolve before you build on this. The repository metadata has an empty licence field, and there is no licence file among the top-level entries. That is not the same as permissive. If you intend to redistribute the Skill, vendor it into a product, or rely on it in a commercial workflow, ask the authors what the terms are, because silence in a repository is a legal question rather than an answer.

The English documentation is a separate file, README.en.md, alongside the Chinese README.md, which is the primary one.

The README opens by promoting a paid product

Above the description of the Skill itself, the README invites readers to try a commercial research product at notyet.chat. The pitch is that you give it an idea and it judges from social platform user feedback whether there is demand, where the pain is, and why users would buy. That is a different product from the open-source Skill, and it is placed first, before the project description.

Two further notes on how the project asks for things. One line asks for a free star on the repository if the project seems useful. Another set of badges points at trendshift.io, and the social links go to a Telegram channel, a MaterialYou Telegram channel and a Discord server.

None of this is hidden, and pointing people at a commercial product from an open-source repository is common. It does mean you should read the Skill's own scope separately from what the maintainers sell, so you know which of the described capabilities are in the open repository and which are in the paid one. The README attributes the eight-platform connected list and the workflow to the Skill itself; it does not say the commercial product is required to use it.

For a project with no releases and a fixed repository history, that distinction is the one to keep straight.

The last push was in April, and there are no releases

The repository has 1,278 stars, 109 forks and 9 open issues, and the last push to main was on 2026-04-27. There are no GitHub releases and no version tags, so there is no way to pin a version of this Skill from the repository side. Whatever you install, you are installing whatever main held on the day you copied it.

That matters more here than for a library, because the platform list is exactly the kind of thing that changes without a version bump. A provider that works this month may be rate limited next month, and the eight-name connected line in the README would still read the same. With no CHANGELOG and no release history, you have no signal about when the list last changed.

What is genuinely well documented is the intent, and it is a coherent one. Rather than building brittle scrapers and per-platform login flows, you ask each vendor's own assistant for the content that vendor can reach, you route the question to the platform that suits it, and you hand the merged result to an agent as structured input. Copy SKILL.md and ROUTING.md into your own setup, keep your own copy of the platform list, and check it against reality periodically.

Editorial conclusion

AI Search Hub fits research on Chinese social platforms where your own scraper would spend its life on logins, CAPTCHAs and rate limits, since the trade is explicitly delegation rather than better scraping. Two things to check before you rely on it. The support table names eleven platforms while the connected list names eight, and the status column has no values at all, so confirm which providers are actually live in ROUTING.md rather than assuming from the table. And note the licence field is empty with no licence file in the tree, the repository has no releases, and the last push to main was 2026-04-27, so pin what you depend on rather than tracking the Skill.

Frequently asked questions

Official sources

1. Issues
2. minsight-ai-info/AI-Search-Hub on GitHub
3. README