Reading social links

Fetching a social media link yourself returns nothing usable. The share URL is tokenised (xsec_token, app_code_link), the content lives in video and images rather than HTML, and the server answers with an app-download shell or a login wall. On Xiaohongshu that shell is an app page whose <title> is the site's slogan and which holds none of the note.

This tool resolves the link, fetches the media with a warmed session, transcribes the audio, describes and OCRs every image, and hands back text.

When to use it

Reach for it the moment a task involves a link you cannot read:

• The user pastes a Xiaohongshu, Douyin, TikTok, YouTube or X link — with or without instructions. A bare link is a request to read it.
• A task references a post you need the contents of: "summarise this", "what does this creator claim", "pull the steps out of this video"
• You already tried fetching a URL and got an app-download page, a login wall or a title with no body

Do not use it as a general web scraper. For an ordinary article that fetches fine, read it directly — this costs credits and takes longer.

Calling it

digest_url(url: "https://v.douyin.com/…")

url is the only required argument. The others:

• format — markdown (default, best for reading) or json (structured fields)
• job_id — collect a digest already running; see below
• breakdown — true adds a teardown of how the post is built (爆款拆解): the hook, timed beats, title formula, call to action and a reusable template. Use it when the user wants to learn from or remake a post. +1 credit
• translate_to — a language code (en, ja, zh-CN): adds a translation beside the original. +1 credit
• ask — a question to answer from the post ("what products and prices does it recommend?", "does the speaker say X?"). The answer comes with the quotes it rests on, each checked against the post; found is false when the post does not say. Use it when the user asked something specific about the link. +1 credit
• partial_ok — true reads the opening minutes of a video that is too long for the plan, instead of refusing it

Long videos need two calls

A short post or a cached link comes back immediately. A long video cannot finish inside one tool call, so the result names a job id instead. When that happens, call the tool again with that job_id and no url — re-sending the url starts the work over and charges for it twice.

digest_url(job_id: "abc123…")

Wait about 15 seconds between attempts. A Xiaohongshu note with images takes one to two minutes; a YouTube video with captions about two and a half.

What comes back

platform, author, title, posted_at, caption, transcript ({t, text}; per line for platform captions, per ~170 s segment for speech recognition), ocr_text, on_screen ({t, text}, approximate times), images ({description, ocr} per image), key_points, stats, tags, source_url, transcript_source, degraded — plus breakdown and translation when asked for.

Two of those change how you should use the result:

• transcript_source is native_captions, asr, gemini_video (an uncaptioned YouTube video watched by Gemini) or none. Captions the platform published are exact; speech recognition is not. If you are about to quote a number or a name back to the user as fact, and the source was asr, say where it came from.
• degraded lists anything that did not fully work, in plain words. It also carries route notes ("read via oEmbed", "read by Gemini watching the video directly") on complete digests. When an entry says something could not be read or failed, the digest is thinner than the post — do not present it as complete.

Platforms

Works: Xiaohongshu (小红书), Douyin (抖音), TikTok, YouTube, X, ordinary web pages.

Not supported: Bilibili returns HTTP 412 to the service's address. Instagram and Facebook need a logged-in session for most posts. If the user asks for one of these, say so rather than trying it.

Cost

10 free credits to start, once per account, no card. A typical post is one credit and video adds one per started minute. breakdown: true (the hook, timed beats and a reusable template — 爆款拆解) and translate_to add one credit each. Anything anyone has digested before is served from cache — free, and back in about a second.