StudyTube

Turn a YouTube topic into a fast study tool: search and tag the best videos, get an AI summary/notes/quiz for any one video, or synthesize a full multi-module course from a topic's videos - each module linked back to the exact video and timestamp that explains it.

Features

• Discovery - search a topic, see ~20 ranked videos with metadata, auto-generated topic tags, and sort/filter (views, length, captions).
• Study - paste a video URL/ID (or pick one from Discovery) to get an AI summary, key points/terms, and a scored 8–10 question quiz, grounded only in that video's transcript.
• Course - turn a topic's videos into a 10–12 module course with per-module overviews, key ideas, checkpoint quizzes, and "watch the source" links that open the right video at the right timestamp.

Setup

1. Python 3.11+, then install dependencies:

   python3 -m venv .venv
   source .venv/bin/activate
   pip install -r requirements.txt

2. Get a YouTube Data API v3 key: in the Google Cloud Console, create a project, enable "YouTube Data API v3" under APIs & Services, then create an API key under Credentials.

3. Get an Anthropic API key: create one at console.anthropic.com.

4. Add both keys to .streamlit/secrets.toml (already gitignored). Copy the template and fill it in:

   cp .streamlit/secrets.toml.example .streamlit/secrets.toml

   YOUTUBE_API_KEY = "..."
   ANTHROPIC_API_KEY = "..."

5. Run it:

   streamlit run app.py

Hosting caveat: transcripts on Streamlit Community Cloud

YouTube blocks transcript requests from cloud-provider IP ranges (AWS/GCP/Azure — which Streamlit Community Cloud runs on). Search and Claude calls work fine hosted; only transcript fetching (and therefore Study/Course) is affected.

• Running locally → everything works, no proxy needed.
• Hosted → add a rotating residential proxy (the app supports Webshare out of the box) by setting WEBSHARE_PROXY_USERNAME / WEBSHARE_PROXY_PASSWORD in the deployed app's Secrets. Without them, search and tagging still work, but Study/Course will show a clear "transcript unavailable" message instead of crashing.

To deploy: push this repo to GitHub (without secrets.toml — confirm .gitignore excludes it), connect it at share.streamlit.io, and paste your keys into the app's Secrets box (same key = "value" TOML format as secrets.toml).

Project layout

app.py                  # entry point: sidebar nav + view routing via session_state
services/
  youtube.py            # search() + video metadata via YouTube Data API
  transcript.py         # get_transcript() / get_transcript_with_timestamps(), proxy-aware
  llm.py                # Anthropic client, call_llm(), safe_json(), long-transcript condensing
  tagging.py            # per-video topic tags (chapters + LLM)
  study.py              # study_materials(transcript) -> summary/notes/quiz
  course.py             # build_course(topic, videos) -> 10-12 module course (map-reduce)
components/
  video_card.py          # one search-result card
  quiz.py                 # interactive scored quiz, shared by Study and Course
  course_view.py          # module rail + progress + checkpoint quizzes + source links

Notes

• Search results, transcripts, study materials, tags, and courses are all cached (@st.cache_data), so re-running a search or revisiting a module never re-spends API quota or re-bills the LLM.
• The course-building flagship feature needs at least 3 videos with usable transcripts; if fewer are available it reports not_enough_transcripts instead of guessing.
• There's no database — session state + caching is enough. Streamlit Community Cloud's filesystem is ephemeral anyway, so nothing here relies on writing to disk.