caishi-mcp

Read Chinese platforms from any MCP client — through your own logged-in Chrome.

Bilibili, Douyin (TikTok China), Xiaohongshu (Little Red Book) and any web page, as plain JSON: titles, captions, authors, engagement stats, media links. Runs on the browser you already use, with the sessions you are already signed in to.

Part of caishi — a reading toolkit for Chinese platforms.

What you get

A self-contained MCP server (the protocol layer is implemented in this package, no MCP SDK dependency). Four tools, one per site command:

| Tool | Args | Returns | |---|---|---| | bilibili/video | bvid (BV id, video URL or b23.tv short link), optional page | title, author, duration, views/likes/coins/favorites, publish date, description | | douyin/detail | target (share short link, video URL or numeric id) | caption, author, digg/comment/collect/share counts, created time, music, no-watermark play address, image gallery | | xiaohongshu/note | url (full note link, ideally with xsec_token) | title, author, full text, likes/collects/comments, tags | | web/read | url, optional wait, selector | any page's main text, through your logged-in browser |

Works from Claude Code, Claude Desktop, Codex, Cursor and any other MCP client.

Requirements

• Node.js >= 20.18.1
• Google Chrome (used in the background alongside your normal browsing; nothing is uploaded)
• An AI client that speaks MCP

Install

npm install -g caishi-mcp

Add it to your MCP client. For Claude Code:

claude mcp add caishi -- npx caishi-mcp

Or in an mcpServers config block:

{
  "mcpServers": {
    "caishi": {
      "command": "npx",
      "args": ["caishi-mcp"]
    }
  }
}

First tool call opens the browser in the background. The engine (OpenCLI) is installed automatically as a dependency. After that, every tool call goes through your real browser — same cookies, same login, same network.

For pages that require a signed-in session (Douyin video pages, Xiaohongshu notes), sign in to that site in Chrome once. The tools then read as you.

Example

Ask your AI client:

Read this Bilibili video and tell me its title, author and view count: https://www.bilibili.com/video/BV...

The client calls bilibili/video with the URL, gets the metadata back, and answers. The same works for a Douyin share link or a Xiaohongshu note.

Notes and limits

• Read-only. No write, publish or account operations are included in this package.
• Rate limits and bot checks of the underlying sites apply. The tools report a page's login wall or risk-control block honestly instead of guessing.
• Xiaohongshu note links work best with a full URL from the site (containing xsec_token).

License

MIT for the code in this package. The caishi project layers content under CC BY-NC-ND 4.0 — see the repository's LICENSE for the full scope.

Credits

Browser engine by OpenCLI (the @jackwener/opencli package). The MCP protocol layer and the site adapters in this package are self-implemented.