这是一个面向内容提取场景的 MCP 服务，目标很直接：

• 你给它一个抖音链接
• 或者给它一个小红书链接
• 它返回结构化信息，并落盘生成脚本文件

当前支持：

• 抖音视频
• 小红书视频笔记
• 小红书图文笔记

默认产物：

• script.md
• info.json

这个项目主要解决四件事：

1. 统一解析抖音和小红书分享链接
2. 自动识别视频笔记还是图文笔记
3. 用云端模型完成视频转写、图片读字、轻量整理
4. 统一输出适合后续内容加工的文件

• 不再只支持抖音，已经扩展到小红书
• 小红书不仅支持视频，也支持图文笔记
• 输出不是一段临时文本，而是固定的 script.md + info.json
• 默认走百炼的轻量模型组合，成本低、速度快、配置简单
• 兼容旧接口，适合接到 Agent Reach 之类的 AI 工作流里

• 想把短视频或图文笔记转成文字稿的内容创作者
• 想把 Douyin / Xiaohongshu 接入自己 AI 工作流的开发者
• 想让 AI 自动读取链接、提取结构化内容、继续写脚本或做分析的人

默认建议直接使用百炼：

• paraformer-v2：视频语音转文字
• qwen3-vl-flash：小红书图文图片读字
• qwen-flash：轻量整理，只做分段、标点、明显错字修正

这套默认组合的目标不是“最强”，而是：

• 简单
• 好用
• 快
• 成本低

script.md 用来给人看，也方便继续交给 AI 做后续处理。通常包含：

• 整理后的脚本
• 原始转写
• 小红书图文的图片文字提取结果

info.json 用来给程序或 Agent 看。通常包含：

• 平台
• 内容类型
• 标题
• 作者
• 发布时间
• 媒体地址
• 采用的 provider 和 model
• 任务状态
• 错误信息

先看这两个 MCP 工具：

• parse_social_post_info
• extract_social_post_script

前者只解析信息，后者直接生成 script.md 和 info.json。

AI-first MCP for extracting structured content from:

• Douyin videos
• Xiaohongshu video notes
• Xiaohongshu image notes

Default outputs:

• script.md
• info.json

This repository is intended to be handed to an AI agent. Write prompts assuming the AI will read this file and perform the setup.

Input:

• one Douyin share link
• one Xiaohongshu share link
• share text that contains one of those links

Output:

• parsed metadata
• raw transcript for videos
• OCR text for Xiaohongshu image notes
• lightly cleaned readable script
• artifacts written to disk

Supported MCP tools:

• parse_social_post_info
• extract_social_post_script
• parse_douyin_video_info
• get_douyin_download_link
• extract_douyin_text

If you are an AI agent configuring this repo, use this default stack unless the user explicitly asks for something else:

• ASR_PROVIDER=bailian
• ASR_MODEL=paraformer-v2
• VISION_PROVIDER=bailian
• VISION_MODEL=qwen3-vl-flash
• CLEAN_PROVIDER=bailian
• CLEAN_MODEL=qwen-flash

Reason:

• paraformer-v2 is the cheapest and fastest default for video ASR
• qwen3-vl-flash is a good default for Xiaohongshu image-note OCR
• qwen-flash is enough for light cleanup: punctuation, paragraphing, obvious typo fixes

Do not switch to heavier models by default.

If you are an AI agent, execute these steps in order:

git clone https://github.com/JNHFlow21/social-post-extractor-mcp.git
cd social-post-extractor-mcp
uv sync

Optional local verification:

python3 -m unittest discover -s tests

Run the MCP server:

uv run python -m social_post_extractor_mcp

Copy the template:

cp .env.example .env

Use this exact default configuration:

ASR_PROVIDER=bailian
ASR_MODEL=paraformer-v2
VISION_PROVIDER=bailian
VISION_MODEL=qwen3-vl-flash
CLEAN_PROVIDER=bailian
CLEAN_MODEL=qwen-flash
BAILIAN_API_KEY=your_bailian_api_key

Do not commit .env.

If you are wiring this repo into an MCP client, use a local config like this:

{
  "mcpServers": {
    "douyin": {
      "command": "/bin/zsh",
      "args": [
        "-lc",
        "cd '/absolute/path/to/social-post-extractor-mcp' && exec '.venv/bin/python' -m social_post_extractor_mcp"
      ],
      "env": {
        "ASR_PROVIDER": "bailian",
        "ASR_MODEL": "paraformer-v2",
        "VISION_PROVIDER": "bailian",
        "VISION_MODEL": "qwen3-vl-flash",
        "CLEAN_PROVIDER": "bailian",
        "CLEAN_MODEL": "qwen-flash",
        "BAILIAN_API_KEY": "YOUR_BAILIAN_API_KEY"
      }
    }
  }
}

Use the shell wrapper form above if your MCP client launches stdio servers from a directory outside this repo. It avoids Python module resolution failures.

If the user is using Agent Reach, the effective system config is usually:

• ~/.mcporter/mcporter.json

Do not assume ~/.agent-reach/tools/.../config/mcporter.json is the active config file.

Check the active source with:

mcporter config list

If mcporter config list shows the douyin server coming from ~/.mcporter/mcporter.json, edit that file instead.

This repo currently assumes Alibaba Cloud Bailian / DashScope for the default path.

Official product page:

• https://www.aliyun.com/product/bailian

Official pricing page:

• https://help.aliyun.com/zh/model-studio/model-pricing

Official first API call / API key doc:

• https://help.aliyun.com/zh/model-studio/getting-started/first-api-call-to-qwen

If you are an AI agent, the purchase flow is:

1. Ask the user to log in to Alibaba Cloud.
2. Tell the user to open the Bailian product page.
3. Tell the user to activate Bailian if it is not activated yet.
4. Tell the user to recharge their Alibaba Cloud account balance.
5. Tell the user to open Bailian console 密钥管理.
6. Tell the user to create an API key.
7. Put that key into local config as BAILIAN_API_KEY.

Current official product page also advertises:

• a one-click API onboarding entry
• free token quota for new users

Important:

• Keep the API key in local config or local environment variables only.
• Do not commit the real key to Git.
• Do not paste the real key into tracked files.

All numbers below are from Alibaba Cloud official pages and should be treated as the current default reference for this repo.

paraformer-v2:

• 0.00008 元 / 秒
• 36,000 秒 monthly free quota shown on the pricing page, equal to 10 小时

fun-asr:

• 0.00022 元 / 秒

Interpretation:

• fun-asr is about 2.75x the price of paraformer-v2
• use paraformer-v2 as the default unless the user explicitly needs a more expensive ASR model

qwen3-vl-flash under the current pricing page tier 0 < Token <= 32K:

• input: 0.15 元 / 百万 Token
• output: 1.5 元 / 百万 Token

qwen-flash under the current pricing page tier 0 < Token <= 128K:

• input: 0.15 元 / 百万 Token
• output: 1.5 元 / 百万 Token

These are rough operating estimates for the default stack.

Using paraformer-v2 only:

• 1 minute video ≈ 0.0048 元
• 3 minute video ≈ 0.0144 元
• 5 minute video ≈ 0.024 元
• 10 minute video ≈ 0.048 元
• 100 videos x 1 minute ≈ 0.48 元
• 100 videos x 3 minutes ≈ 1.44 元
• 100 hours total ≈ 28.8 元

For qwen-flash, cleanup is usually negligible compared with ASR.

Reference estimate:

• assume 2,000 input tokens + 2,000 output tokens for one short transcript
• estimated cleanup cost ≈ 0.0033 元 / 条

Formula:

• input cost = input_tokens / 1,000,000 * 0.15
• output cost = output_tokens / 1,000,000 * 1.5

For qwen3-vl-flash, OCR cost depends on:

• image count
• image size
• prompt tokens
• OCR output length

Use this as the operating rule:

• image-note OCR is still cheap for normal creator workflows
• if the user mainly processes videos, budget primarily by ASR
• if the user mainly processes long multi-image notes, monitor token usage from actual runs instead of guessing

Code handles:

• link parsing
• Douyin / Xiaohongshu detection
• note type detection
• metadata extraction
• artifact directory creation
• writing script.md and info.json

Cloud models handle:

• paraformer-v2: video speech-to-text
• qwen3-vl-flash: image text extraction
• qwen-flash: light readability cleanup

This is not "the LLM figures everything out by itself". The MCP does the workflow orchestration; the models only handle recognition and light cleanup.

If you are an AI agent using this repo:

1. Prefer extract_social_post_script over platform-specific tools.
2. Keep extract_douyin_text only for backward compatibility.
3. Keep cleanup light. Do not summarize unless the user asks.
4. Preserve raw transcript and raw OCR text in artifacts.
5. Do not change the default model stack unless the user asks.
6. Do not store real API keys in tracked files.

Run these after setup:

python3 -m unittest discover -s tests

Example MCP smoke test:

mcporter call 'douyin.extract_social_post_script(share_link: "https://v.douyin.com/xxxxx/", output_dir: "/tmp/social-post-extract")'

• AGENT_REACH_INTEGRATION.md

Apache 2.0. See LICENSE.