Eleven days ago I open-sourced auto-xhs, a system that lets a cloud AI agent run a social account end to end: research, writing, video, posting, and following up on comments. It works through the browser like a person would — no unofficial APIs, no bulk-engagement tricks.
Last week I wrote about the first external user, who ran the pipeline, measured the output videos frame by frame, and filed three real bugs our validator had missed. That story ended with a promise: if he sent a PR, his would be reviewed first and used first. The fixes are now in, and issue #1 is closed.
Since then the repo has grown in three ways:
Five platforms, one chain. The same daily chain now covers Xiaohongshu, Douyin/TikTok, X, YouTube, and WeChat Official Accounts. Each platform has its own module with its own rules and runnable scripts: subtitle burning and clip splicing for Douyin, a post checker for X, an upload checker for YouTube, a draft builder for WeChat. The Xiaohongshu chain is still the most complete; the others are honest about what they do and do not automate.
English by default. All docs and script output are now in English, with a Chinese version linked at the top of every file. And the first thing a new agent does after reading the repo is ask the operator: Chinese or English? The whole session then follows that choice. No more guessing from file names.
A first release. v0.1.0 is out. The honest numbers today: 13 stars, 1 fork.
What has not changed: the compliance rules. No bulk interaction, no reposting other people's content, human pace. And a few things are deliberately not automated: posting to X itself (anti-ban), uploading to YouTube itself (use the official API), and the final WeChat publish (manual). The repo says so in plain words instead of pretending otherwise.
If you run it and your videos look wrong in a way the validator misses, that is still the report I want most. And if you have content that repeats daily — weather, a digest, a ranking — playbooks live in scenarios/, and new ones are welcome. The collection thread is issue #2.
Top comments (1)
Five platforms in 11 days is an incredibly fast pace.
The smartest part of this update is the deliberate limitation: not automating the final publish on X or YouTube. Headless bots get banned instantly on those platforms, so keeping the final trigger manual is a sign you actually understand the compliance landscape rather than just brute-forcing it with a browser script.
Since it works through the browser like a person, how are you handling session persistence? Are you using something like Playwright with saved auth states/cookies, or does the operator have to manually log in for every run?
Keep up the great work. 13 stars is how every big open-source project starts!