One workspace that summarizes and compares websites, YouTube videos, documents, images, audio, and video — powered by a local Python extraction pipeline and LiteLLM model routing.
Status: beta desktop/web AI workspace. The repo name still reflects the original scraper/Discord-bot roots, but the current product is AI Media Studio.
| What it is | A multimodal AI chat workspace: paste a URL, YouTube link, or drop a file, and get an AI summary/analysis — with per-chat memory. |
| Install | Download a release installer, or run from source (see Quick Start). |
| Inputs | Websites · YouTube · text · PDF/DOCX/PPTX/XLSX/RTF/MD/CSV/JSON/HTML/XML · images · audio · video. |
| Requires | Python 3.10+, an AI provider key in .env; for full features: ffmpeg, tesseract, Playwright Chromium. |
| Models | 22+ models across 5 providers (OpenAI, Anthropic, Google, Together/Llama, MiniMax) — or any LiteLLM string. Set TEXT_AI_MODEL. |
| Why it's robust | Layered extraction with fallback chains (esp. YouTube transcripts) so a summary still works when one path is blocked. |
| Drag & drop | Drop files straight onto the desktop window to analyze them. |
| Local-first | DB, downloads, transcripts, and uploads stay in user-writable local locations. You pay your own provider API usage. |
| Honesty layer | The app carries which path actually succeeded into the summary, to avoid faking that a fallback worked. |
Who it's for: anyone who wants a single tool to read/watch/listen for them — researchers comparing a paper against an article, people summarizing long videos, or anyone tired of juggling separate scraper / transcript / OCR / vision tools.
Jump to: Screenshots · Quick Start · Architecture · What it can do · Supported inputs & file types · Extraction pipeline · Models · Desktop install · Run from source · Build installers · Config defaults · Troubleshooting
- Handles many input types in one workspace instead of separate tools.
- Uses extraction fallbacks so summaries can still work when one source path fails.
- Keeps per-chat memory and uploaded context local to the app workflow.
- Supports multiple AI providers through LiteLLM.
- Summaries are only as reliable as the extracted source content and the selected model.
- Some websites, videos, or files may block extraction or require fallback analysis.
- API keys and local dependencies are required for full multimodal functionality.
Sidebar chat management, hero banner, chat area with an AI-summarized research paper, and the input bar with file attachment.
A YouTube link is pasted in chat; the AI extracts the transcript, summarizes key points, and answers follow-ups.
A website is scraped and summarized, then a PDF is uploaded and compared against the article — multi-step analysis in one chat.
Get the latest from the Releases page:
| Platform | File |
|---|---|
| macOS (Apple Silicon — M1/M2/M3/M4) | AI-Media-Studio-1.0.0-macOS-Apple-Silicon.dmg |
| macOS (Intel) | AI-Media-Studio-1.0.0-macOS-Intel.dmg |
| Windows (64-bit) | AI-Media-Studio-Setup-1.0.0-Windows-x64.exe |
See Installation (Desktop App) for prerequisites and first-launch steps.
git clone https://github.com/Evan1108-Coder/Website-Youtube-File-AI-Scraper.git
cd Website-Youtube-File-AI-Scraper
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# Edit .env with the AI provider keys you want to use
uvicorn src.ai_scraper_bot.webapp:app --reloadThe web server runs at http://127.0.0.1:8000 (or port 18919 when launched via Electron).
Everything from the web version, plus:
- Native app — launches from Applications/Start Menu, with its own dock icon.
- Drag-and-drop — drop files directly onto the window to analyze them.
- Auto-starts backend — no need to manually run a Python server.
- Hidden inset titlebar (macOS) — clean, modern appearance.
- Persistent storage — database and downloads in user-writable locations.
- Normal AI chat with memory per chat.
- Multiple chats in a left sidebar (limit 10), with rename / clear / delete / clear-all controls.
- Upload files directly in the browser or by drag-and-drop.
- Summarize websites and YouTube links.
- Analyze documents, images, audio, and video.
- Per-chat drafts & attachment state, a pause button while the AI works, local SQLite history, and a banner hide/show toggle.
It's designed to feel like a normal AI chat screen, with the project's multimodal extraction pipeline behind it.
Inputs: websites · YouTube links · text questions · uploaded documents / images / audio / video.
| Category | Extensions |
|---|---|
| Text & markup | .txt .md .csv .json .html .xml |
| Documents | .pdf .docx .pptx .xlsx .rtf |
| Images | .png .jpg .jpeg .avif |
| Audio | .mp3 .wav .m4a .aac .flac .ogg |
| Video | .mp4 .mov |
The project avoids dead-end failures by using layered extraction.
Page-text extraction → related useful-URL collection → image review when relevant → directly-downloadable website-video review when possible → a final summary focused on the subject matter, not just page structure.
- optional
YouTube Data APImetadata youtube-transcript-apiyt-dlpsubtitle attemptDownSub + PlaywrightSaveSubs + Playwright- metadata fallback
So the app can still produce something useful even if direct YouTube access is partially blocked.
The active visual-description path uses your configured AI model — image descriptions and video key-frame descriptions both come from your model (the older BLIP caption path is no longer the normal active flow).
The media pipeline separates transcript/speech analysis, visual analysis, and music analysis — so a silent video can still be reviewed visually, and a music-heavy file can still produce music analysis even if speech transcription is weak.
Essentia— default music feature layer (BPM, key, loudness-like values).AcoustID— optional song ID via local fingerprinting + API key.MIRFLEX— optional repo hook for future music tagging/classification.
If one music stage fails, the others continue.
The app carries extra runtime/extraction context into the summary pipeline — which YouTube path succeeded, which music libraries were attempted vs produced output, recent runtime diary lines, and which media was actually reviewed — to reduce fake claims like "a fallback worked" when it didn't.
Set TEXT_AI_MODEL in your .env to any of these (or any LiteLLM-compatible model string):
| Provider | Models |
|---|---|
| OpenAI | gpt-5.5-pro, gpt-5.5, gpt-5.5-mini, gpt-5.4-pro, gpt-5.4-mini, gpt-4o, gpt-4o-mini |
| Anthropic | claude-opus-4-7, claude-sonnet-4-7, claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5, claude-3.5-sonnet |
gemini-3.1-pro, gemini-3-flash, gemini-2.5-flash-lite |
|
| Together AI (Llama) | llama-4-maverick, llama-4-scout, llama-3.3-70b |
| MiniMax | minimax-m3, minimax-m2.7, minimax-m2.5 |
- Download the
.dmgfor your chip (Apple Silicon or Intel). - Open the DMG and drag AI Media Studio to Applications.
- On first launch, right-click the app → Open (to bypass Gatekeeper — the app is unsigned).
- The app auto-starts the Python backend.
- Download the
.exeinstaller. - Run it and follow the prompts.
- Launch AI Media Studio from the Start Menu.
- Python 3.10+ in PATH, with
pip install -r requirements.txt. - System tools:
ffmpeg,tesseract(audio/OCR features). - Playwright Chromium:
playwright install chromium(website scraping). .envwith your AI API keys (see.env.example).
- Database:
~/Library/Application Support/ai-media-studio/webapp.sqlite(macOS) /%APPDATA%/ai-media-studio/webapp.sqlite(Windows). - Downloads:
~/Documents/AI Media Studio Downloads/.
Full guide in SETUP.md.
pip install -r requirements.txt
npm install
npm start # Electron app in dev mode
# or run just the web server:
PYTHONPATH=src python -m ai_scraper_bot.webapp| Area | Files |
|---|---|
| Web entrypoint | src/ai_scraper_bot/webapp.py |
| Web backend | src/ai_scraper_bot/web/service.py, web/store.py |
| Web frontend | src/ai_scraper_bot/web/static/index.html, app.css, app.js |
| Shared config / prompts | src/ai_scraper_bot/config.py, prompts.py |
| Summarizer (LiteLLM) | src/ai_scraper_bot/services/summarizer.py |
| YouTube extraction | src/ai_scraper_bot/services/youtube.py |
| Website extraction | src/ai_scraper_bot/services/website.py |
| Transcript-site fallbacks | services/downsub.py, services/savesubs.py |
| Transcription | src/ai_scraper_bot/services/transcription.py |
| Local video / vision / music | services/video_analysis.py, services/vision.py, services/music_analysis.py |
| File parsing | src/ai_scraper_bot/parsers/file_parser.py |
npm run build:mac # macOS (arm64 + x64)
npm run build:win # Windows (x64)
npm run build:all # both platformsOutput goes to dist-electron/.
ENABLE_LOCAL_VISION=true
ENABLE_MUSIC_DETECTION=true
MUSIC_ESSENTIA_ENABLED=true
MUSIC_ACOUSTID_ENABLED=false # until AcoustID is configured
MUSIC_MIRFLEX_ENABLED=false # until MIRFLEX is actually set up
YOUTUBE_COOKIE_MODE_ENABLED=false
YOUTUBE_DOWNSUB_ENABLED=true
YOUTUBE_SAVESUBS_ENABLED=true
YOUTUBE_TRANSCRIPT_SITE_HEADLESS=trueFull env reference: ENVREADME.md.
Web UI ↔ FastAPI server ↔ specialized extractors (web, YouTube, documents, vision, audio, music), with LLM calls routed through LiteLLM to multiple providers.
How URLs, YouTube links, documents, images, audio, and video flow through layered extraction with fallback chains, converging into a final LiteLLM summary stored in per-chat SQLite memory.
- API keys belong in
.envand should never be committed. - Uploaded files, scraped content, transcripts, and summaries may be stored locally by the app.
- Content can be sent to the configured AI provider when you ask the app to summarize/analyze it.
- Don't upload private/sensitive files unless you understand your local setup and selected provider.
- Before sharing the repo, do not expose
.env,.venv, downloaded test media, cookies files, browser profile exports, real API keys, or machine-specific secrets (.gitignorecovers these — double-check anyway). - See SECURITY.md for vulnerability reporting and secret handling.
YouTube still fails — that doesn't mean the whole app is broken; check which stage failed (youtube-transcript-api, yt-dlp, DownSub, SaveSubs, metadata fallback).
Audio transcription fails with NumPy / torch issues — the local Whisper setup is currently intended to run with numpy<2.
A video has no audio — some .mp4/.mov files are silent: transcript audio analysis won't run, but visual review still can; music analysis only runs if an audio stream exists.
MIRFLEX enabled but not working — it's an optional repo hook; the rest of the music chain continues regardless.
More: TROUBLESHOOTING.md.
| Topic | Document |
|---|---|
| Full setup & configuration | SETUP.md |
| Every environment variable | ENVREADME.md |
| Troubleshooting | TROUBLESHOOTING.md |
| Security policy | SECURITY.md |
| Contributing | CONTRIBUTING.md |
MIT License — © Evan Lu. See LICENSE.
These visuals are generated from the actual repository structure and project workflow, not placeholders.
