Video Digest Agent

Automatically monitors YouTube channels, YouTube keyword searches, and Bilibili user spaces for new videos, extracts transcripts, auto-detects the content type (stocks, crypto, podcast, tech review, education, news, etc.), generates tailored summaries using your choice of LLM (Gemini, OpenAI, or Claude), and delivers them via email, local file, or both.

Includes a web UI for point-and-click configuration and one-click runs, as well as a CLI for automation and scripting.

How It Works

The app uses an agent-based pipeline to produce accurate, content-aware summaries:

1. DISCOVER  — Fetches new uploads from YouTube channels, search results, or Bilibili users
2. CLASSIFY  — Detects video type (stock analysis? podcast? tutorial?)
3. PROMPT    — Selects the best summary template for that content type
4. SUMMARIZE — Generates a structured summary via LLM
5. VERIFY    — (Optional) Second LLM pass to catch errors & hallucinations

Supported Content Types

| Type | What you get | |------|-------------| | Stock Analysis | Sentiment score, tickers, bull/bear thesis, price targets, conviction levels | | Macro Economics | Economic outlook, indicators, central bank commentary, sector views | | Crypto | Market sentiment, token analysis, on-chain signals | | Podcast/Interview | Guest bios, discussion points, notable quotes, contrarian views | | Tech Review | Specs, pros/cons, comparisons, verdict | | Educational | Core concepts, step-by-step process, pitfalls, key takeaways | | News | Key facts, perspectives, implications | | General | Auto-structured summary adapted to content |

Setup

1. Get API Keys

YouTube Data API v3 Key (required for YouTube sources)

1. Go to Google Cloud Console
2. Create a new project (or select an existing one)
3. Go to APIs & Services > Library
4. Search for "YouTube Data API v3" and click Enable
5. Go to APIs & Services > Credentials
6. Click Create Credentials > API Key
7. Copy the key — this is your YOUTUBE_API_KEY
8. (Recommended) Click Restrict Key and limit it to YouTube Data API v3 only

LLM API Key (choose one)

| Provider | Cost | How to get the key | |----------|------|--------------------| | Gemini (recommended) | Free tier | Go to Google AI Studio → Create API Key | | OpenAI | ~$0.02-0.10/video | Go to OpenAI Platform → Create new secret key | | Anthropic (Claude) | ~$0.02-0.06/video | Go to Anthropic Console → API Keys → Create Key |

Gmail App Password (only needed if OUTPUT_MODE is email or both)

1. Go to Google Account Security
2. Enable 2-Step Verification if not already on
3. Go to App Passwords
4. Select Mail and your device, then click Generate
5. Copy the 16-character password — this is your SENDER_PASSWORD

Note: If you don't use Gmail, update SMTP_SERVER and SMTP_PORT for your provider (e.g., Outlook: smtp.office365.com:587).

Tip: If you just want to save summaries locally without email, set OUTPUT_MODE=local and skip this step entirely.

2. Install Dependencies

# Required for Whisper audio fallback (Mac)
brew install ffmpeg

pip3 install -r requirements.txt

Note: ffmpeg is only needed if a video has no captions and Whisper kicks in. The app tries YouTube captions first (instant), and only downloads + transcribes audio as a fallback.

3. Configure Environment

cp .env.example .env

Edit .env with your actual values:

# YouTube channels to monitor (comma-separated, without @)
YOUTUBE_CHANNELS=RhinoFinance,MeetKevin

YOUTUBE_API_KEY=AIza...

# Pick your LLM: gemini, openai, or anthropic
LLM_PROVIDER=gemini
GEMINI_API_KEY=AIza...

# Summary languages (up to 2)
SUMMARY_LANGUAGES=English,Chinese

# Verify accuracy with a second LLM pass (optional)
VERIFY_SUMMARY=false

# Output: email, local, or both
OUTPUT_MODE=email

# Email (only needed when OUTPUT_MODE is email or both)
SENDER_EMAIL=you@gmail.com
SENDER_PASSWORD=abcd efgh ijkl mnop
RECIPIENT_EMAILS=you@gmail.com

Usage

Web UI

python3 app.py

macOS — port 5000 already in use? macOS Monterey and later runs AirPlay Receiver on port 5000 by default. Either disable it (System Settings → General → AirDrop & Handoff → AirPlay Receiver → off) or use a different port:

python3 app.py --port 8080

Opens a dashboard at http://127.0.0.1:5000 with:

• Dashboard — live stats, recent history, config overview
• Run — trigger once, poll, test a specific video, retry failed, or validate config
• Config — edit all settings through a form (no manual .env editing required)
• Archive — browse and read saved summary files

python3 app.py --port 8080 --host 0.0.0.0   # custom port / expose to network

CLI

Check once for new videos

python3 main.py

Run continuously (checks every hour)

python3 main.py --poll

Test with a specific video

# YouTube video
python3 main.py --video dQw4w9WgXcQ

# Bilibili video (prefix BV IDs are detected automatically)
python3 main.py --video BV1xx411c7XZ

Dry run (no email sent — prints summary to stdout)

python3 main.py --video dQw4w9WgXcQ --dry-run

Validate your configuration

python3 main.py --check

Show processing history

python3 main.py --history

Retry previously failed videos

python3 main.py --retry

Video Sources

YouTube Channels

Set YOUTUBE_CHANNELS to a comma-separated list of channel handles (without @). The agent resolves handles to channel IDs and caches them locally to minimise API usage.

YouTube Keyword Search

Set YOUTUBE_SEARCH_QUERIES to discover new videos beyond your subscribed channels:

YOUTUBE_SEARCH_QUERIES=AI news,machine learning,LLM
YOUTUBE_SEARCH_MAX_RESULTS=5          # results per query
YOUTUBE_SEARCH_INTERVAL=14400         # search every 4 hours (seconds)
YOUTUBE_SEARCH_QUOTA_BUDGET=5000      # max API units/day for search
YOUTUBE_SEARCH_RELEVANCE_KEYWORDS=AI,LLM,GPT   # title pre-filter
YOUTUBE_SEARCH_MIN_DURATION=10        # skip clips shorter than N minutes
YOUTUBE_SEARCH_MAX_TOTAL=15           # cap total videos per search cycle
YOUTUBE_SEARCH_MIN_VIEWS=1000         # skip low-traffic videos (0 = off)

Keyword search uses the YouTube Search API (100 units/call). The daily free tier is 10,000 units total across all API calls.

Bilibili

Monitor Bilibili user spaces for new uploads. Requires browser cookies for subtitle/transcript access:

BILIBILI_ENABLED=true
BILIBILI_USERS=12345,67890            # numeric UIDs from profile URLs
BILIBILI_SESSDATA=...                 # get from browser DevTools
BILIBILI_BILI_JCT=...
BILIBILI_BUVID3=...

To get the cookies: open bilibili.com, log in, open DevTools (F12) → Application → Cookies, and copy the values for SESSDATA, bili_jct, and buvid3.

Bilibili support requires the optional bilibili-api-python and httpx packages (included in requirements.txt).

Configuration Reference

Core

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | YOUTUBE_CHANNELS | One source required | — | Comma-separated channel handles (without @) | | YOUTUBE_API_KEY | If YouTube enabled | — | YouTube Data API v3 key | | LLM_PROVIDER | No | gemini | LLM to use: gemini, openai, or anthropic | | GEMINI_API_KEY | If gemini | — | Google Gemini API key | | GEMINI_MODEL | No | gemini-3.1-pro-preview | Gemini model to use | | GEMINI_FALLBACK_MODELS | No | see .env.example | Fallback model chain when primary hits quota | | OPENAI_API_KEY | If openai | — | OpenAI API key | | OPENAI_MODEL | No | gpt-4o-mini | OpenAI model to use | | ANTHROPIC_API_KEY | If anthropic | — | Anthropic API key | | ANTHROPIC_MODEL | No | claude-sonnet-4-5-20250929 | Claude model to use | | SUMMARY_LANGUAGES | No | English | Up to 2 languages, comma-separated | | VERIFY_SUMMARY | No | false | Enable accuracy verification pass | | OUTPUT_MODE | No | email | email, local (save file only), or both | | SMTP_SERVER | No | smtp.gmail.com | SMTP server | | SMTP_PORT | No | 587 | SMTP port | | SENDER_EMAIL | If email/both | — | Email to send from | | SENDER_PASSWORD | If email/both | — | SMTP password / app password | | RECIPIENT_EMAILS | If email/both | — | Email(s) to send summaries to (comma-separated) | | POLL_INTERVAL | No | 3600 | Seconds between checks |

YouTube Search

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | YOUTUBE_SEARCH_QUERIES | No | — | Comma-separated search terms; leave empty to disable | | YOUTUBE_SEARCH_MAX_RESULTS | No | 5 | Results per search query (1–50) | | YOUTUBE_SEARCH_INTERVAL | No | 14400 | Seconds between search runs | | YOUTUBE_SEARCH_QUOTA_BUDGET | No | 5000 | Max YouTube API units for search per day | | YOUTUBE_SEARCH_RELEVANCE_KEYWORDS | No | see .env.example | Title pre-filter keywords | | YOUTUBE_SEARCH_MIN_DURATION | No | 10 | Skip videos shorter than N minutes | | YOUTUBE_SEARCH_MAX_TOTAL | No | 15 | Max total search results processed per cycle | | YOUTUBE_SEARCH_MIN_VIEWS | No | 1000 | Skip videos with fewer views (0 = off) |

Bilibili

| Variable | Required | Default | Description | |----------|----------|---------|-------------| | BILIBILI_ENABLED | No | false | Enable Bilibili monitoring | | BILIBILI_USERS | If enabled | — | Comma-separated numeric UIDs | | BILIBILI_SESSDATA | If enabled | —