Our own visit statistics are always on. Google Analytics, which helps us understand how people use our public pages (not your signed-in workspace), is enabled only with your permission. See our Privacy policy.

RedNote & Douyin MCP Server on ModelScope: Remote, No Login

Add EveryInfra from the ModelScope MCP plaza and let your AI read Xiaohongshu (RedNote) notes and Douyin videos. Real timings, prices, config and errors.

Co-branded card: EveryInfra × ModelScope, Now on the ModelScope MCP hub

To give your AI client live Xiaohongshu (RedNote) and Douyin data, add the EveryInfra entry from the ModelScope MCP plaza and put your EveryInfra API key in the request header. The server runs remotely. You do not open a local browser, and you do not log in to RedNote or Douyin. You pay per call, and failed calls and empty results are free.

We connected with the config from the ModelScope entry in the early hours of October 22, 2026 (Singapore time). The tool list came back in 0.21 seconds without a key. One page of RedNote search results took 3.56 seconds and cost $0.005556. One Douyin video detail took 17.62 seconds and cost $0.000556. This guide shows which entry to pick, how to configure it, what each call costs and what the errors mean.

Three entries: which one to add

Since October 21, the ModelScope MCP plaza (China site, modelscope.cn) lists three EveryInfra entries, all of type "Hosted":

EntryModelScope pagePick it when
EveryInfra: Xiaohongshu, Douyin, Taobao and JD dataxiaoyuaimaomao/EveryInfraYou need several platforms, or web search and CAPTCHA solving too
Xiaohongshu data MCPxiaoyuaimaomao/everyinfra-xiaohongshuYou only work with RedNote: notes, comments, creators, topics, trending
Douyin data MCPxiaoyuaimaomao/everyinfra-douyinYou only work with Douyin: videos, transcripts, comments, creators, products

All three connect to the same server and use the same key. The single-platform pages describe only one platform, which helps the AI pick the right capability. The main entry covers all 91 platforms and 426 capabilities, plus 17 search tools and 52 CAPTCHA types.

In one sentence: **the EveryInfra ModelScope entry is a remote MCP server; your AI calls one tool, everyinfra_call_api, with a platform and a capability name, gets structured public data, and each call is charged to your EveryInfra balance.**

Remote server or local server?

Search for Xiaohongshu (RedNote) in the ModelScope MCP plaza and you find many similar entries. Their pages describe two common designs. Some run on your computer and need you to log in to RedNote in a local browser. Others, like EveryInfra, run remotely. Both work; they fit different jobs:

DesignWhat you prepareAccount exposureGood fit
Local server with your own platform loginNode or Python, a local browser, a logged-in sessionUses your own accountReading your own messages or saves, posting, liking
EveryInfra remote serverOne EveryInfra API keyNo platform accountReading public data: search, detail, comments, profiles, trending, products

EveryInfra only reads public data. It does not post, like or read private messages. If you need "post this note for me", choose a local server.

Illustration comparing two platforms. Left: a laptop next to a local server box, with a browser open on screen and a personal account card with a lock and a profile icon plugged into it; a shield with an exclamation mark sits on the card, and several cables tangle between laptop, browser and server. Right: the same laptop holds only a small key, and one teal line runs up to a remote server in a cloud, which reads a few public photo and short-video cards and sends them back as a neat stack of data blocks.
A local server needs a browser and your own platform login. The remote server needs one EveryInfra API key and reads only public data.

Test 1: what you see after connecting

At 02:19 on 2026-10-22 (UTC+8) we sent tools/list to https://api.everyinfra.com/mcp without a key:

  • 25 tools came back in 0.21 seconds. The mainland China endpoint api-cn.everyinfra.com/mcp returned the same 25 tools in 0.22 seconds.
  • The three tools you use most: everyinfra_list_capabilities (free: platforms, parameters, prices), everyinfra_call_api (run one data capability) and everyinfra_search (17 search tools). The others manage the account, usage, billing, keys and support tickets.

Test 2: RedNote search and a Douyin video

Same session, with a test key, one call each through everyinfra_call_api:

CallParametersTimeResultCharge
RedNote keyword searchxiaohongshu / search, keyword "camping gear" (sent in Chinese)3.56 s20 notes this time (a page holds at most 20), with a next_page_token; likes from 2 to 22,790; posted 2026-01-06 to 2026-10-21; 2 video notes$0.005556
Douyin video detaildouyin / video, one MediaStorm video17.62 sLikes 99,689, comments 6,458, shares 21,357, saves 8,875; 57 s long; view_count 0$0.000556

What we learned:

  1. Counts are live. On October 21 the same Douyin video had 99,673 likes and 21,352 shares. One day later: 99,689 and 21,357. Store the read time with every number.
  2. **Ignore Douyin view_count.** Douyin does not publish view counts. The field is usually missing or 0, and it does not mean zero views. Use likes, comments and shares (field notes in the Douyin scraper API docs). Bilibili and TikTok do return real views.
  3. RedNote search returns one page per call. A page holds at most 20 notes and limit has no effect. For more, send the next_page_token back as page_token; each page is one charged call. The token is valid for 24 hours.
Screenshot of one real request and response: POST /api/v1/social with platform douyin, action video and the link to the MediaStorm video. The response is 200 after 16.75 seconds with a charge of $0.000556. The result shows author MediaStorm, 57 seconds long, 100401 likes, 6476 comments, 21471 shares, 8930 saves and view_count 0.
The raw request and response (excerpt) for the same Douyin video, read again on October 25 at 20:02 (UTC+8): likes have grown since October 22, and `view_count` is still 0. Douyin video API

Test 3: errors cost nothing

We then made two mistakes on purpose with the test key. Neither call was charged:

CaseResponseHow the AI recovers
Misspelled parameter sorttunknown_param: "sortt is not a parameter of xiaohongshu/search — did you mean sort?", with required keyword and optional content_type, date_range, limit, page_token, sort; "Nothing was charged."Rename to sort and resend
No keyunauthorized: "no API key provided", with a sign-up link; "Nothing was charged."Ask the user for a key

Both come back as normal MCP tool results with isError: true, so the AI can read the reason and the next step. Full error list: API error handling.

Connect in 3 steps

  1. Get a key. Sign up at everyinfra.com and create an API key in the console (quickstart).
  2. Copy the config. Open one of the entries and paste its config into your client (Claude Desktop, Cursor, Cherry Studio or any client with remote MCP). The type value depends on the client: streamable_http, streamableHttp or http.
json
{
  "mcpServers": {
    "everyinfra": {
      "type": "streamable_http",
      "url": "https://api.everyinfra.com/mcp",
      "headers": { "Authorization": "Bearer <your EveryInfra API key>" }
    }
  }
}
  1. Ask a free question first. Ask "Use EveryInfra to list what RedNote data you can get." The AI calls everyinfra_list_capabilities, which is free. Then ask a real question, such as "Search RedNote for notes about camping gear, sort by likes, and list title, likes and link."

A note on Qoder CN (Alibaba's IDE assistant in China): its docs say the IDE's MCP plaza shows recommended servers plus every ModelScope MCP server, that the IDE cannot search the plaza yet, and that it connects up to 10 servers at once (Qoder CN docs, in Chinese). If one-click install does not work there, add the server by hand with the config above.

Screenshot of the Alibaba Cloud Help Center page on configuring MCP servers in Qoder CN (in Chinese), section "Option 1: add from the MCP plaza". The teal boxes mark two sentences: the IDE cannot search the MCP plaza yet, and the MCP plaza tab shows recommended servers plus every MCP server from the ModelScope community. On the right is the page's own screenshot of the IDE with the plaza list and one-click install.
Qoder CN docs (Chinese): the IDE's MCP plaza shows every ModelScope MCP server and cannot be searched yet. The teal boxes are the two sentences cited here. Alibaba Cloud Help Center: configure MCP servers (Chinese original)

Common capabilities and prices

PlatformCapability (action)PriceReference
RedNoteKeyword search search (at most 20 per page)$0.005556 per pageRedNote search API
RedNoteNote detail note$0.005556 per callRedNote note API
RedNoteTrending trending$0.005556 per callRedNote scraper API
DouyinVideo detail video$0.000556 per callDouyin video API
DouyinTranscript transcript$0.000556 per callDouyin transcript API
TikTokKeyword search search$0.000556 per callTikTok search API
JDAll SKUs and prices skus$0.001389 per rowJD SKU API

Prices come from the live capability catalog, read on 2026-10-22 at 02:18 (UTC+8). The pricing page is the reference. Data starts at $0.56 per 1,000 calls.

Cost example: a weekly China content scan

A brand team that tracks Chinese consumer trends runs this every week:

  • 50 keywords, one RedNote page each (1,000 notes): 50 × $0.005556 = $0.28
  • 200 Douyin video details: 200 × $0.000556 = $0.11
  • Total: about $0.39 a week. Failed calls and empty results are free.

To read comments on 50 of those notes, see RedNote comments in batches; the first page of each costs $0.005556.

Same capability from code

MCP sits on a normal HTTP API, so scripts can call it directly:

bash
curl -X POST https://api.everyinfra.com/api/v1/social \
  -H "Authorization: Bearer $EVERYINFRA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"platform":"xiaohongshu","action":"search","params":{"keyword":"camping gear"}}'

Fields you will use most:

FieldMeaning
id, urlNote ID and link; the link usually carries an access token, so a logged-out browser can open it
title, textTitle and a text excerpt (about 60 characters in search; call note for the full text)
like_count, collect_count, comment_count, share_countPublic counts at read time; null means not returned, not zero
posted_atPost time, ISO 8601, UTC
author_name, author_idAuthor name and 24-character user ID; pass the ID to profile
is_video, duration_secVideo note flag and length
next_page_tokenCursor for the next page; empty means no more pages

The response also has billing.charged_usd (the charge for this call) and pagination.has_more. General remote MCP setup: remote MCP guide. Platform pages: RedNote MCP, Douyin MCP, all entry points at MCP server. Plugin and skill files are in our GitHub repository.

Recent changes (September–October 2026)

  • Sep 23: Qoder CN updated its MCP docs. The IDE's MCP plaza is the recommended list plus every ModelScope MCP server (Alibaba Cloud docs, in Chinese). A ModelScope listing is also browsable inside Qoder CN.
  • Sep 29: ASUS added an MCP Marketplace to AI SuperBuild. It integrates ModelScope natively and lets developers browse more than 1,000 MCP tools and connect them to an on-device model (ASUS blog). ModelScope entries now reach hardware vendors' clients.
  • Sep 15: Cloudflare split bot controls into Search, Training and Agent (Cloudflare blog). Sites now treat an agent that reads a page for a person differently from training crawlers, which matters to anyone feeding live web data to an AI.
  • Oct 20: QuestMobile's Q3 report. China's AI office tools had 142 million active users in September, up 50.6% year on year (36Kr summary). These agents need outside data, and MCP plazas are where users find it.

FAQ

Do I need a RedNote or Douyin account? No. You put only your EveryInfra API key in the header. The server runs remotely, so no local browser is needed.

What does "Hosted" on the ModelScope page mean? It is the type ModelScope shows for the entry. Our server runs at api.everyinfra.com; the entry gives you the config that points there.

Can a single-platform entry call other platforms? All three entries use the same server and key, so technically yes. The single-platform pages describe only one platform, which keeps the AI focused. Use the main entry for cross-platform work.

Related reading

Reference LF-18