Video Fetch — YouTube & Bilibili

Fetch, transcribe, and summarize YouTube or Bilibili videos using subtitles, cloud STT, local Whisper, or description fallback with proxy and login support.

Video Content Fetcher (YouTube & Bilibili)

Before You Start

This skill needs 1-3 things configured depending on your use case:

What	Why	How
Residential IP proxy	YouTube blocks datacenter IPs. You MUST use a residential broadband proxy (e.g., HK home broadband). VPS/cloud IPs will fail.	--proxy socks5h://127.0.0.1:1080 (also install pysocks: pip install pysocks)
ElevenLabs API key	Default STT engine for audio transcription when no subtitles exist. Far more accurate than local Whisper, especially for Chinese.	export ELEVENLABS_API_KEY="sk_..." or --stt-api-key @~/.elevenlabs_key. Get key: https://elevenlabs.io/app/settings/api-keys
Bilibili cookie	Bilibili AI subtitles require login. Without cookie, falls back to STT.	Export SESSDATA + bili_jct from browser, save to ~/.bilibili_cookie, use --cookie @~/.bilibili_cookie. See details below.

Minimum setup for YouTube: proxy + ElevenLabs API key (or local Whisper) Minimum setup for Bilibili: ElevenLabs API key (or local Whisper). Cookie optional but recommended.

Quick Start

bash

# YouTube
python3 scripts/youtube_fetch.py "https://www.youtube.com/watch?v=VIDEO_ID" --proxy socks5h://127.0.0.1:1080

# Bilibili
python3 scripts/youtube_fetch.py "https://www.bilibili.com/video/BVxxxxxxxxxx"

Fallback Chain

text

1. Subtitle/Transcript API  (highest fidelity — full spoken content)
       |  if unavailable
       v
2. ElevenLabs STT  (default, cloud-based, high accuracy with punctuation)
       |  if no API key or fails
       v
3. Local Whisper  (auto-fallback, requires local install)
       |  if unavailable
       v
4. Video description  (lowest fidelity — creator's written summary only)

Always tell the user which source was used. If description-only, explicitly note that it's not a full transcript.

Platform Support

Platform	Subtitle source	STT fallback	Description fallback
YouTube	youtube-transcript-api	ElevenLabs / Whisper	HTML meta extraction
Bilibili	Player API (cookie required for AI subs)	ElevenLabs / Whisper	View API

Platform is auto-detected from URL. Force with --platform youtube or --platform bilibili.

Usage

bash

# YouTube with proxy
python3 scripts/youtube_fetch.py VIDEO_URL --proxy socks5h://127.0.0.1:1080

# Bilibili (cookie required for subtitles — without it, falls back to STT/description)
# IMPORTANT: always use @filepath to avoid exposing credentials in shell history / process list
python3 scripts/youtube_fetch.py "https://www.bilibili.com/video/BVxxxxxxxxxx" --cookie @~/.bilibili_cookie

# Bilibili short link
python3 scripts/youtube_fetch.py "https://b23.tv/xxxxxx"

# Specify transcript languages (default: zh-Hans,zh-Hant,zh,en)
python3 scripts/youtube_fetch.py VIDEO_URL --langs "en,ja" --proxy PROXY

# Force local Whisper instead of ElevenLabs
python3 scripts/youtube_fetch.py VIDEO_URL --stt whisper --whisper-model medium

# Disable audio transcription entirely
python3 scripts/youtube_fetch.py VIDEO_URL --stt none

# JSON output
python3 scripts/youtube_fetch.py VIDEO_URL --json --proxy PROXY

# Save to file
python3 scripts/youtube_fetch.py VIDEO_URL --output /tmp/transcript.txt --proxy PROXY

Dependencies

Required:

• youtube-transcript-api (pip) — YouTube transcript fetching
• requests (pip) — HTTP requests for APIs

Required for STT (audio transcription):

• yt-dlp (pip or brew) — audio download from YouTube/Bilibili

Optional (for local Whisper, used as auto-fallback when ElevenLabs is unavailable):

• openai-whisper (pip) — local audio transcription
• ffmpeg (system) — required by whisper

Optional (for SOCKS5 proxy):

• pysocks (pip) — required if using socks5h:// proxies

bash

# Required
pip install youtube-transcript-api requests

# Required for STT
pip install yt-dlp

# If using SOCKS5 proxy
pip install pysocks

# Optional: local Whisper fallback
pip install openai-whisper
brew install ffmpeg  # or apt install ffmpeg

ElevenLabs STT (Default)

ElevenLabs Scribe v2 is the default speech-to-text engine. It provides significantly better accuracy than local Whisper, especially for Chinese — correct characters, automatic punctuation, and proper formatting (e.g., book title marks).

Setup:

bash

# Option 1: Environment variable (recommended)
export ELEVENLABS_API_KEY="sk_your_key_here"

# Option 2: Key file
echo "sk_your_key_here" > ~/.elevenlabs_key && chmod 600 ~/.elevenlabs_key
# Then use: --stt-api-key @~/.elevenlabs_key

If no API key is available, the skill automatically falls back to local Whisper.

Security: Same rules as Bilibili cookie — use env var or @filepath, never pass the key directly on the command line.

Get your API key at: https://elevenlabs.io/app/settings/api-keys

Proxy Setup

YouTube requires a proxy from geo-blocked regions. Bilibili typically works directly from China.

• SOCKS5: socks5h://127.0.0.1:1080
• HTTP: http://127.0.0.1:1081

Important: Use residential IP proxies for YouTube. YouTube blocks datacenter/cloud server IPs. Residential broadband IPs (e.g., Hong Kong home broadband) work reliably, while VPS/cloud IPs will likely be blocked. Recommend routing only YouTube domains through the proxy.

Bilibili Cookie (Required for Subtitles)

Bilibili subtitles are almost entirely AI-generated and require login to access. Without cookies, the subtitle API returns empty results for virtually all videos, and the skill will fall back to STT or description.

How to get your cookie:

1. Log in to bilibili.com in your browser
2. Open DevTools (F12) > Application > Cookies
3. Copy the values of SESSDATA and bili_jct
4. Save to a file: echo "SESSDATA=xxx;bili_jct=xxx" > ~/.bilibili_cookie && chmod 600 ~/.bilibili_cookie
5. Use: --cookie @~/.bilibili_cookie

Security warning:

• Your Bilibili cookie is a login credential. If leaked, others can act as you (post comments, send messages, etc.)
• Always use --cookie @filepath instead of passing the cookie string directly on the command line, because:
  • Command-line arguments are visible in ps aux to other users on the same machine
  • They get saved to shell history (~/.zsh_history, ~/.bash_history)
  • They may appear in Claude Code tool call logs
• Set file permissions: chmod 600 ~/.bilibili_cookie
• Add to .gitignore if the file is anywhere near a git repo
• Cookies expire periodically — refresh from browser when needed

Translation & Language

Respond in the user's conversation language. Do not default to any hardcoded language.

• If the user speaks English, summarize in English
• If the user speaks Chinese, summarize in Chinese
• If the user speaks Japanese, summarize in Japanese
• If the transcript language differs from the user's language, translate while summarizing

This requires no extra configuration — Claude naturally adapts to the conversation language.

When Summarizing

• Transcript/subtitle source: Summarize freely — this is the complete spoken content
• ElevenLabs STT source: High accuracy with punctuation — treat as near-transcript quality
• Whisper STT source: Reliable but may have character-level errors in Chinese — note this to the user
• Description source: Warn the user that this is the creator's written summary, not the full video content
• For long transcripts (>50K chars), break into sections and summarize incrementally