Video Download Transcribe

多平台视频下载 + 本地转录 + 视频内容分析。 **触发词**：这个视频说了什么、视频内容是什么、帮我看这个视频、下载这个视频、视频转录、字幕提取、B站视频、抖音视频、bilibili、youtube视频、帮我转录 **支持平台**：B站/抖音/TikTok/YouTube/小红书/微博/快手

Security Scan

The skill largely does what it says (download + local transcription) but contains undeclared external API hooks, hard-coded local paths, and install/runtime steps that can send URLs/data to third-party services — these mismatches warrant caution.

Purpose & Capability

Name/description match the included tools (ffmpeg, yt-dlp, Playwright, local Whisper). However the packaged server code references multiple cloud services (SILICONFLOW, MINIMAX, THIRD_PARTY_API/TikHub) and default third-party endpoints that are not declared in the skill metadata — these are not strictly required for the stated core purpose and should have been documented as optional/required credentials.

Instruction Scope

SKILL.md tells the agent/user to run setup.sh and to call local MCP endpoints (douyin-analyzer) and remote MCP (tikhub). The server code and instructions will: run subprocessed Python Playwright scripts, call third-party APIs to fetch video info/CDN URLs, and may fall back to cloud STT. The instructions do not clearly warn that video URLs or content may be sent to external endpoints (e.g., liuxingw.com, tikhub). The agent is therefore directed to perform network calls and remote API usage beyond simple local download/transcription without declaring those data flows.

Install Mechanism

There is no official install spec (instruction-only), but a provided setup.sh will attempt to install Playwright/Chromium using an npm mirror and depends on an openclaw-media venv being present. The script assumes specific local paths and may fail or require manual steps. The install approach is not high-risk by itself, but it relies on external mirrors and repositories and hard-coded venv locations.

Credentials

Registry lists no required env vars, yet SKILL.md and server.py reference several environment variables (DOUYIN_CHROMIUM_PATH, DOUYIN_STT_MODE, SILICONFLOW_API_KEY, MINIMAX_API_KEY, WHISPER_MODEL_DIR, TRANSCRIPTS_DIR, DOUYIN_THIRD_PARTY_API). Cloud STT or image analysis will use API keys and default external base URLs; those credentials are not declared in metadata and could be requested by the code at runtime. The skill also uses default third-party endpoints (e.g., liuxingw.com) which could receive video/share URLs.

Persistence & Privilege

always is false and the skill does not request permanent platform-wide privileges. setup.sh and server.py create local directories and expect to run as a local MCP service, but there is no evidence the skill tries to modify other skills or enable itself globally.

What to consider before installing

This skill's core idea (download video, run local transcription) is coherent, but the included code and instructions will — by default or on failure — call external services and use environment keys that are not declared in the registry. Before installing or running: 1) Inspect server.py to see which external domains (THIRD_PARTY_API, SILICONFLOW_BASE_URL, MINIMAX_BASE_URL, the default liuxingw.com) will receive requests; 2) Do not provide API keys (SILICONFLOW_API_KEY, MINIMAX_API_KEY, TikHub token) unless you trust those services; 3) Prefer local-only mode (ensure DOUYIN_STT_MODE=local and install/models locally) if you want transcripts to stay on your machine; 4) Be aware setup.sh references hard-coded user paths (e.g., /Users/kk) and expects an existing openclaw-media venv — run it in a sandbox or review/adjust the script before execution; 5) If you must use the skill, run it in an isolated account/container and monitor outbound network traffic to confirm no unexpected data exfiltration.

多平台视频下载 + 本地转录

功能特点

• ✅ 纯浏览器下载（免费） — 抖音用 Playwright 无水印抓取，无需 API Key
• ✅ MLX Whisper — Apple Silicon 原生 Metal GPU 加速（Mac M系列首选）
• ✅ faster-whisper — CPU/CUDA 通用兜底（非 Apple Silicon 也能用）
• ✅ 短视频同步转录 — <5 分钟视频直接返回完整结果
• ✅ 长视频后台转录 — ≥5 分钟自动后台运行，返回 transcript_id 可查询
• ✅ 可检索 — 转录结果支持关键词查询相关片段 + 时间戳

支持的平台

平台	检测关键词	下载方式
B站	bilibili.com/video, b23.tv, BV号	yt-dlp
抖音	douyin.com, v.douyin.com	Playwright 浏览器抓取 或 TikHub
TikTok	tiktok.com	TikHub
YouTube	youtube.com, youtu.be	yt-dlp
小红书	xiaohongshu.com, xhslink.com	TikHub
微博	weibo.com, m.weibo.cn	TikHub
快手	kuaishou.com, ksurl.cn	TikHub

---

🔑 核心原则

所有平台统一两步走

获取下载链接 → curl/yt-dlp 下载 → mlx_whisper 转录

原因：

• analyze_video 把下载+转录绑在一起，长视频容易超时
• 两步走每步都 <10秒，稳定可控
• 每步可单独排错

排错核心原则

当一个工具失败时，不要放弃！找替代方案。

---

初始化安装（首次使用）

抖音 Playwright Chromium 安装

# 运行 setup.sh 自动安装
cd ~/.openclaw/workspace/skills/video-download-transcribe/douyin-mcp/
./setup.sh

setup.sh 会：

1. 检测系统环境
2. 查找或安装 Chromium（使用国内 npmmirror 镜像）
3. 设置 DOUYIN_CHROMIUM_PATH 环境变量
4. 测试浏览器抓取是否正常

手动安装（如果 setup.sh 失败）：

# 方式 A：用 openclaw-media 的 playwright 安装 chromium
git clone https://github.com/openclaw/openclaw-media ~/openclaw-media 2>/dev/null || true
PLAYWRIGHT_DOWNLOAD_HOST=https://npmmirror.com/mirrors/playwright \
  ~/openclaw-media/.venv/bin/playwright install chromium

# 方式 B：自行安装 playwright
pip install playwright
PLAYWRIGHT_DOWNLOAD_HOST=https://npmmirror.com/mirrors/playwright playwright install chromium

环境变量（运行 AI 助手前设置）：

export DOUYIN_CHROMIUM_PATH="$HOME/Library/Caches/ms-playwright/chromium-1105/chrome-mac/Chromium.app/Contents/MacOS/Chromium"

---

工作流程（统一两步走）

第一步：获取下载链接

# ========== 抖音 ==========

# 方式 A：douyin-analyzer 浏览器抓取（免费，需 Chromium）
mcporter call douyin-analyzer.get_douyin_download_link share_link:"https://v.douyin.com/xxx/"

# 方式 B：tikhub-douyin 获取实时 CDN 地址（更可靠，需 API Key）
mcporter call tikhub-douyin.douyin_web_fetch_video_high_quality_play_url share_url:"https://v.douyin.com/xxx/"

# ========== B站/YouTube/其他平台 ==========

# 用 yt-dlp 直接下载（会自动处理重定向和 Cookie）
yt-dlp -o "/tmp/video.mp4" "https://www.bilibili.com/video/BVxxx"

# ========== TikTok/小红书/微博/快手 ==========

# tikhub-douyin 获取下载链接
mcporter call tikhub-douyin.douyin_web_fetch_one_video_by_share_url share_url:"<链接>"

第二步：下载 + 转录

# 用 curl 下载（抖音/TikTok 获取的 CDN 地址）
curl -L -o /tmp/video.mp4 "https://cdn.example.com/video.mp4"

# 转录（本地 mlx_whisper）
python3 << 'PYEOF'
import os
os.environ["HF_HUB_OFFLINE"] = "1"
import mlx_whisper
result = mlx_whisper.transcribe(
    "/tmp/video.mp4",
    path_or_hf_repo="mlx-community/whisper-small-mlx",
    verbose=True
)
for seg in result.get("segments", []):
    print(f"[{seg['start']:.1f}s - {seg['end']:.1f}s] {seg['text']}")
PYEOF

长视频（≥5分钟，后台转录）

# 用 analyze_video 获取 transcript_id（后台转录中）
mcporter call douyin-analyzer.analyze_video url:"https://..." --timeout 300000

# 等待 1-2 分钟后获取结果
mcporter call douyin-analyzer.get_transcript transcript_id:"xxx"

---

MCP 工具

douyin-analyzer（本地进程，免费）

# 获取无水印下载链接（浏览器抓取）
mcporter call douyin-analyzer.get_douyin_download_link share_link:"https://v.douyin.com/xxx/"

# 解析视频基本信息
mcporter call douyin-analyzer.parse_douyin_video_info share_link:"https://v.douyin.com/xxx/"

# 通用分析（仅用于获取 transcript_id，后台转录）
mcporter call douyin-analyzer.analyze_video url:"https://..." --timeout 300000

# 获取长视频后台转录结果
mcporter call douyin-analyzer.get_transcript transcript_id:"xxx"

# 检索转录内容
mcporter call douyin-analyzer.query_transcript transcript_id:"xxx" query:"关键词" top_k:3

返回结构（get_douyin_download_link）：

{
  "video_id": "xxx",
  "download_url": "https://www.iesdouyin.com/aweme/v1/play/?video_id=xxx",
  "share_link": "https://v.douyin.com/xxx/"
}

⚠️ download_url 是 iesdouyin.com 重定向 URL，有时效性，建议立即使用。

tikhub-douyin（远程服务，需认证）

# 获取视频信息+播放链接（所有平台通用）
mcporter call tikhub-douyin.douyin_web_fetch_one_video_by_share_url share_url:"<链接>"

# 获取最高画质 CDN 地址（推荐，最可靠）
mcporter call tikhub-douyin.douyin_web_fetch_video_high_quality_play_url share_url:"<链接>"

返回字段：

{
  "data": {
    "original_video_url": "https://cdn.example.com/..."
  }
}

---

❗️ 排错流程

1. 工具返回 error/失败
   ↓
2. 抖音：换一个 MCP（douyin-analyzer ↔ tikhub-douyin）
       B站/YouTube：检查 yt-dlp 是否支持
       TikTok/其他：检查 tikhub-douyin 是否可用
   ↓
3. 抖音：检查 Chromium 是否安装 + DOUYIN_CHROMIUM_PATH 是否设置
   ↓
4. 如果都失败 → 找图文/文字版内容

常见问题

问题	原因	解决
download_url 为空	Playwright Chromium 未安装	运行 setup.sh 或手动安装
DOUYIN_CHROMIUM_PATH 警告	环境变量未设置	设置 export DOUYIN_CHROMIUM_PATH=...
curl 下载失败	抖音 URL 已过期	改用 TikHub 获取实时 CDN
TikHub 返回 401	API 认证失败	检查 mcporter.json 的 TikHub Token
yt-dlp 下载失败（B站）	Cookie/地区限制	加代理或用 TikHub
analyze_video 超时	视频较大	改用两步走

Chromium 安装（国内镜像）

# 检查是否已安装
ls ~/Library/Caches/ms-playwright/

# 国内镜像安装
PLAYWRIGHT_DOWNLOAD_HOST=https://npmmirror.com/mirrors/playwright \
  ~/openclaw-media/.venv/bin/playwright install chromium

---

本地转录（mlx_whisper）

import os
os.environ["HF_HUB_OFFLINE"] = "1"  # 必须放在 import 之前！

import mlx_whisper

result = mlx_whisper.transcribe(
    "/path/to/video.mp4",
    path_or_hf_repo="mlx-community/whisper-small-mlx",
    verbose=True
)

for seg in result.get("segments", []):
    print(f"[{seg['start']:.1f}s - {seg['end']:.1f}s] {seg['text']}")

推荐 Whisper 模型

模型	参数量	速度	精度
whisper-small-mlx	244M	快	中等
whisper-base	74M	中	中
whisper-large-v3-turbo	809M	慢	高

---

踩坑记录

2026-04-08 抖音下载修复

问题：douyin-analyzer 无法获取下载链接

排查过程：

1. 官方 API Cookie 失效（tt_webid=xxx 是假的）
2. 第三方 API (liuxingw.com) 返回 HTML 而非 JSON，已失效
3. TikHub API 需要认证，公共接口返回 401
4. Playwright Chromium 未安装（网络问题导致安装超时）
5. Chromium 版本不匹配（1.58 需要 chromium-1208，镜像只有 1105）
6. subprocess 中的 chromium 路径变量无法传递到代码字符串
7. iesdouyin.com 重定向 URL 有时效性

最终解决方案：

• 使用 openclaw-media 的 playwright 1.42.0（chromium-1105）
• 在 subprocess 代码字符串中硬编码 chromium 路径
• 用 DOUYIN_CHROMIUM_PATH 环境变量配置路径
• 两步走替代 analyze_video（下载+转录分离）

教训：

• 当工具失败时，列出所有 MCP 找替代方案
• MCP 内部变量无法传递到 subprocess 字符串代码
• analyze_video 把下载+转录绑在一起容易超时，两步走更稳定
• TikHub 返回 original_video_url 是真实 CDN 地址，比 iesdouyin.com 重定向 URL 更可靠
• f-string 中的 {{ 会先被外层 f-string 处理，需用 dict() 代替字典字面量