Wechat Reader

A minimal local MCP server that lets Claude Code read public WeChat Official Account articles (mp.weixin.qq.com) and view their images.

What it does

• read_wechat_article(url) — fetches an article via headless Chromium and returns it as markdown with image placeholders and a numbered image list.
• view_wechat_image(url) — downloads one image from mmbiz.qpic.cn and returns it as MCP ImageContent so Claude can actually look at it.

Why local-first

Many existing WeChat scrapers route requests through third-party extraction services, which means the URLs you read and the full article bodies leave your machine. This server runs 100% locally:

• Article extraction uses your own headless Chromium via Playwright.
• Image downloads go directly to mmbiz.qpic.cn with the Referer header omitted to bypass hotlink protection — no third-party image proxy.
• Nothing is persisted: no inbox, no cache, no cookies between calls.

Install

git clone https://github.com/zzysgdhrrr-coder/wechat-reader-mcp.git
cd wechat-reader-mcp

python3 -m venv .venv
.venv/bin/pip install -U pip
.venv/bin/pip install "mcp[cli]>=1.0" "playwright>=1.40" "requests>=2.28" "idna>=3.6" "Pillow>=10.0"
.venv/bin/playwright install chromium

Wire up to Claude Code

Add this to ~/.claude/settings.json under mcpServers (replace /absolute/path/to/wechat-reader-mcp with where you cloned the repo):

{
  "mcpServers": {
    "wechat-reader": {
      "command": "/absolute/path/to/wechat-reader-mcp/.venv/bin/python",
      "args": ["/absolute/path/to/wechat-reader-mcp/server.py"]
    }
  }
}

Restart Claude Code.

Use

In Claude Code:

读这篇文章 https://mp.weixin.qq.com/s/xxx，用三句话总结
读这篇文章 https://mp.weixin.qq.com/s/xxx，把第二张图里的财务数据列出来

Claude will call read_wechat_article first, then view_wechat_image on specific images only when the question requires visual content.

Security model

• Domain whitelist: read_wechat_article only accepts https://mp.weixin.qq.com/s/...; view_wechat_image only accepts *.qpic.cn URLs. Anything else is rejected.
• SSRF protection: hostnames are resolved and any private/loopback/ link-local/cloud-metadata IP is blocked. IDN homograph attacks rejected.
• Process isolation: each article fetch starts a fresh Chromium context, no cookies persist between calls.
• Size caps: article body capped at 500 KB, single image at 5 MB.
• No SVG: SVG image content is rejected (causes Claude Code session crashes — see anthropics/claude-code#28279).
• Token-aware resize: images are downscaled to 2576px long edge before base64 encoding to bound LLM token cost.
• Trust boundary marker: article markdown is wrapped in <wechat_article>...</wechat_article> to remind the LLM the content is external data, not instructions (basic prompt-injection hygiene).

Limits

• Public articles only. Paid / followers-only / 关注后可见 articles cannot be fetched — there is no login flow on purpose, to avoid storing your WeChat session cookie.
• No comments / video. Article body only.
• Hotlink-protected placeholder image is a known failure mode: if WeChat returns the "此图片来自微信公众平台" placeholder instead of the real image, the response will look unusually small (< 5 KB).