I build LinkDigest. Every number below is from a measured run recorded in the project log; nothing is projected.
The problem
Agents are good at reading and bad at opening. Give Claude Code, Cursor, or a Dify workflow a link from Xiaohongshu, Douyin, TikTok, YouTube or X and the usual tool — fetch the URL, parse the HTML — returns one of three things: an app-download shell, a login wall, or a page whose body is 264 characters of navigation because the post itself lives in a JSON blob or in a video. The agent reports "nothing useful here" and it is telling the truth about the HTML.
The substance of these posts is not text. It is speech in a video, words printed on images, screenshots of code, a recipe laid out in a photo. A plain fetch throws all of that away, and a share link from the app carries a token that expires in a few weeks, so even the HTML you do get is different depending on which link and which user-agent you used.
The solution
