energypantry/agent-browser-runtime/xiaohongshu-search

Xiaohongshu (小红书 / RED) Search, Note & Profile

Description

Use Agent Browser Runtime to search Xiaohongshu (小红书 / RED), read note bodies and comment threads, and collect user-profile data through a persistent logged-in Chrome profile. All actions run through the humanized UI primitives (brs ui) and are parsed by the bundled extractors/golddigger-xhs.extract.js (search / note / profile modes).

Use this when a task needs real Chinese social-media UGC: word-of-mouth, product or store reviews, 探店/种草 notes, competitor/creator account research, consumer trends, or "what do commenters actually say" questions. Do NOT use it for English/overseas info, breaking news, technical docs, or papers — those belong to general web search.

Xiaohongshu content is subjective UGC and frequently contains soft ads / sponsorships.

Report sample sizes (how many notes/comments read) and cross-check before strong claims.

Prerequisites

• Agent Browser Runtime stack running from the repo root (docker compose up -d),

extension connected (brs status shows "extensionConnected": true).

• A persistent Chrome profile that is logged in to xiaohongshu.com (log in once via noVNC).
• Python 3 (used only for pretty-printing JSON).

Commands

A wrapper is provided at scripts/xiaohongshu/xhs.sh (run it from anywhere; it resolves the repo root itself). Each command is serial and takes ~15–60s (humanized scrolling + comment expansion). Do not run concurrently.

cd scripts/xiaohongshu
 
# Ensure runtime is up and a xiaohongshu tab is open (caches lease/tab in $TMPDIR)
./xhs.sh setup
 
# Search: prints a numbered note list (noteId / likes / author / title).
# Raw JSON saved to $TMPDIR/xhs_last_search.json
./xhs.sh search "keyword" [N=20]
 
# Read the Nth note from the last search: body + image count + top-level comments + replies.
# Raw JSON saved to $TMPDIR/xhs_last_note.json
./xhs.sh read 1
 
# User profile: nickname, redId (小红书号), bio, IP location, following/followers,
# likes+saves, and recent notes. Pass a userId or a full profile URL.
./xhs.sh profile 5e3d7d02000000000100795b
./xhs.sh profile "https://www.xiaohongshu.com/user/profile/<userId>"
 
# Fallback: navigate directly to a note URL/id. Often returns 404/300031 on a stale
# xsec_token — prefer `read` (clicking a card mints a fresh token).
./xhs.sh open <noteUrl|noteId>

Standard workflow

1. search "<keyword>" to get candidate notes.
2. Pick high-like / relevant notes and read <n> for body + comment threads.
3. For author/competitor research, take the userId/profile URL and run profile <id>.
4. Summarize: distinguish note-author claims vs. comment-section consensus vs. likely ads;

state how many notes/comments were read.

What you can / cannot get

Available: note title, body, hashtags, image count, publish date, author nickname/userId; comment text, commenter nickname, comment likes, nested replies (verified on a live note: 50 top-level + 167 replies = 217 comments with zero missing nicknames); profile: nickname, redId (小红书号), bio, IP location, following/followers, likes+saves count, recent-note list.

Not available (platform limits, not tooling bugs):

• Another user's saved / liked tabs — Xiaohongshu does not render these on others'

profiles (privacy); no tool can read them for an account you don't own.

• Deleted / followers-only / geo-restricted content.
• Direct navigation to a note URL often fails xsec_token validation (404 / 300031):

the token is minted by page JS when a card is clicked. Always open notes via read (card click), never by hand-building URLs.

Maintenance

If Xiaohongshu changes its DOM and fields come back empty, update the selectors in extractors/golddigger-xhs.extract.js against fresh HTML (brs browse-html <lease> <tab>):

• Search cards: section.note-item; title .title a span; author .name; likes .count.
• Note: #detail-title / #detail-desc; comment unit div[id^="comment-"].comment-item

(replies additionally have comment-item-sub); nickname a.name; body div.content.

• Profile: .user-name / .user-desc / .user-redId (redId) / .user-IP; numeric stats

.user-interactions .count in order: following, followers, likes+saves.