Hermes Agent and Python automate Xiaohongshu AI-news publishing
A project author reports building an open-source, unattended Xiaohongshu pipeline that uses a cron-scheduled Hermes Agent to search for AI news, verify source dates, write five-point summaries, render two 1080×1440 cards, publish through browser automation, and confirm public visibility. Because the author says Xiaohongshu has no open publishing API or MCP server, the system uses the Chrome DevTools Protocol to control a persistent Edge instance, with nine Python scripts handling deterministic steps. The author reports that the system ran from September 14 to November 2, 2026, producing 35 publications, three failed task runs, 15 no-run days because the local gateway was down, and no duplicate publications; these figures were not independently verified.
The report presents an open-source example of separating agent judgment—such as selecting and organizing news—from deterministic execution such as uploading, publishing, and verification. It also documents concrete browser-automation problems, including a publish button inside a closed shadow DOM and unreliable editor key events.
从 0 到持续运行：用 Hermes Agent + Python 搭一个小红书 AI 资讯自动发布系统
每天 9 点，一个 Agent 会自己去搜当天的 AI 新闻、逐条打开原文核实日期、写成 5 条要点、渲染成两张 1080×1440 的卡片、上传到小红书创作平台、填标题正文、点发布，然后回头确认这篇笔记到底有没有真的发出去、是不是公开可见。
这套东西从 2026-09-14 跑到 2026-11-02，一共发出 35 篇，中间失败 3 次。
这篇文章想讲清楚两件事：一个没有开放 API 的站点，怎么被自动化到可以无人值守；以及哪些活该交给 Agent，哪些活必须写成确定性脚本。代码全部开源在文末。
一、为什么做这个项目
起因很具体：我每天要看 AI 圈的更新，读完又觉得"不如顺手整理出来"。但手工做这件事的成本不在写，在重复的机械动作——搜、点开、核对日期、排版、上传、填标题、点发布、确认发没发出去。全是可枚举的步骤，一步不能少，而且每天一模一样。
#Hermes Agent#Xiaohongshu#Chrome DevTools Protocol#Microsoft Edge#Python#Cron