Video Analyzer CN

视频内容分析工具。支持B站、抖音、今日头条视频链接。 发送视频URL → 自动下载 → 抽帧 → 本地AI逐帧识别 → 综合总结。 使用本地minicpm-v模型，无需云端API。

Security Scan

The skill largely matches its stated purpose (download → frame → local model analyze) but contains a few inconsistencies and instructions that could access the user's browser or an external 'agent-reach' service; verify those before installing.

Purpose & Capability

The name/description match the included scripts: download (douyin_download.py) and per-frame analysis (analyze_frames.py) submitting base64 images to a local model API. Requested tools (ffmpeg, yt-dlp, Python, Chrome, local minicpm-v/ollama) are consistent with the stated workflow. Minor mismatch: SKILL.md refers to references/analyze.py but the repo has scripts/analyze_frames.py — likely a documentation vs file-layout inconsistency that will break automated runs unless corrected.

Instruction Scope

Instructions tell the agent to extract Douyin video URLs via browser devtools (Chrome MCP) and to possibly use an external 'agent-reach' douyin MCP service. Browser automation means the agent would interact with the user's browser DOM (potentially exposing pages/tokens) — this is sensitive. The skill also instructs manipulating the PATH in a PowerShell snippet and uses hard-coded local temp paths. All network calls in code target video hosts and localhost:11434 (a local model server), but the mention of an external agent-reach MCP is an out-of-band dependency that could route data off-device if used.

Install Mechanism

No install spec (instruction-only plus small Python scripts). Nothing in the manifest downloads or executes remote archives during install. Risk from install-time code is low.

Credentials

The skill requests no environment variables or credentials (good). However it uses hard-coded Windows paths (C:\Users\39535\.openclaw\workspace\tmp and D:\AI\ffmpeg), and assumes a local model API at http://localhost:11434 — these are plausible but user-specific assumptions. The browser-based extraction step could access browser state; SKILL.md explicitly warns not to use cookies-from-browser due to Chrome cookie encryption, but the agent still needs browser access to retrieve video.src. No secrets are requested by the skill itself.

Persistence & Privilege

always is false and there are no service/account modifications. The skill does not request permanent platform-level privileges. It writes temporary files to a workspace tmp path (documented) and expects the user/agent to clean them up.

What to consider before installing

What to check before installing: - Correctness: SKILL.md refers to references/analyze.py but the provided script is scripts/analyze_frames.py — confirm filenames and paths so the agent will actually run the analyzer. - Local model: The analyzer sends base64 images to http://localhost:11434/api/generate (common Ollama default). Ensure you run and trust a local model server on that port before using the skill; otherwise the requests will fail or hit an unexpected service. - Browser automation: The skill asks the agent to extract video.src from pages using Chrome devtools (MCP). That requires the agent to interact with your browser; consider whether you trust the agent to access your open browser tabs and DOM. Prefer manually supplying the direct video URL if you are uncomfortable. - External services: The doc mentions an optional 'agent-reach' douyin MCP service. Avoid using any external MCP service unless you understand where data (video URLs or frames) will be sent — that could leak video content or metadata off your machine. - Test with non-sensitive content first: Run the skill on a short public video to confirm behavior, temp file locations, and that only localhost and the video hosts are contacted. - Clean up: Confirm the temporary workspace path and delete temp videos/frames after use (the docs describe cleanup but verify it runs). If the author provides corrected SKILL.md (pointing to the actual analyzer file) and clarifies that no external agent-reach service is required (or documents exactly when it's used and where it runs), this assessment could be upgraded to 'benign'.

Video Analyzer — 视频内容分析

支持平台

平台	URL特征	下载方式
B站	b23.tv / bilibili.com	yt-dlp 直接下载
抖音	v.douyin.com / douyin.com	浏览器提取URL → Python带Referer下载
今日头条	m.toutiao.com / toutiao.com	yt-dlp 指定格式下载

完整流程

URL → 识别平台 → 下载视频 → ffmpeg抽帧 → minicpm-v逐帧识别 → 综合总结

使用方式

用户发送视频链接，agent 自动执行以下步骤：

Step 1: 识别平台并下载

读取 references/download.md 获取各平台下载脚本。

Step 2: 抽帧

# 设置ffmpeg路径
$env:Path = "D:\AI\ffmpeg;$env:Path"

# 抽帧（每N秒一帧，根据视频长度调整）
# 短视频(<60s): fps=1/3 (每3秒一帧)
# 中视频(1-5min): fps=1/5 (每5秒一帧)  
# 长视频(>5min): fps=1/10 (每10秒一帧)
ffmpeg -i <video.mp4> -vf "fps=1/5,scale=640:-1" -q:v 3 <output_dir>/frame_%03d.jpg -y

关键参数：

• scale=640:-1 缩小分辨率，原图会超时
• -q:v 3 JPEG质量（1最好，31最差，3够用）
• 帧数控制在 5-30 帧之间最佳

Step 3: 逐帧识别

读取 references/analyze.py 脚本，用 minicpm-v 逐帧识别。

Step 4: 综合总结

收集所有帧的描述，综合输出视频内容总结。

环境依赖

工具	路径	用途
ffmpeg	D:\AI\ffmpeg\	视频抽帧
yt-dlp	系统PATH	B站/头条下载
minicpm-v	ollama本地	图片识别
Python	系统PATH	抖音下载+分析脚本
Chrome	系统安装	抖音视频URL提取

临时文件

所有临时文件存放在 C:\Users\39535\.openclaw\workspace\tmp\：

• 视频文件: *.mp4
• 抽帧图片: frame_*.jpg / dy_frame_*.jpg / tt_frame_*.jpg
• 分析脚本: analyze_*.py / douyin_download*.py

分析完成后可清理临时文件。

注意事项

• 抖音需要Chrome浏览器打开页面提取视频URL（需浏览器MCP）
• 抖音视频URL有时效性，提取后需立即下载
• Chrome v20 cookies加密问题，不能用 --cookies-from-browser chrome
• 长视频帧数多时处理时间较长（每帧约20-30秒）
• 内存紧张时优先清理其他进程