Snag：把收藏链接和文件变成可执行笔记的捕获应用

Save anything. Get the idea back.

Snag is a capture app with one job: turn anything you save into something you can actually use. Send it a link or a file and it comes back with a summary, the key ideas, why it matters, and a triage that tells you what to do next. Everything lands in a searchable vault, so nothing you save dies in a folder.

Two surfaces drive one pipeline: a Telegram bot and a local web app.

The loop

Save, extract, understand, recommend.

1. Save. Send a link (TikTok, YouTube, X, Instagram, Facebook, any article) or upload a file.
2. Extract. Video is transcribed locally with faster-whisper. YouTube uses native captions. Text is pulled straight off the page. There is no paid speech-to-text anywhere in the pipeline.
3. Understand. DeepSeek writes the note: summary, key ideas, why it worked, why it matters, a reusable pattern, recommendations, and tags. A second pass adds a triage: stage, action type, impact, and effort. Real view, like, and save counts pulled from the source ground the impact score, so it reflects proven reach instead of a guess.
4. Recommend. Two to four concrete next actions, saved to the vault with the full transcript.

Why it exists

Saved folders are where ideas go to die. Snag forces a decision on every save: act on it, reference it, or file it. The name is the product. Snag it all.

Architecture

Three tiers, one pipeline.

• service.py: the pure capture pipeline, zero Telegram imports. capture_url, capture_file, and capture_text each return a CaptureResult.
• bot.py: a thin Telegram adapter over the service.
• webview/ and web/: a localhost web API and a TanStack Start frontend.

Supporting modules: ingest.py (yt-dlp plus ScrapeCreators and ScrapTik failover), social_capture.py (Playwright for Instagram and Facebook), analyze.py (faster-whisper plus DeepSeek), and db.py (SQLite vault with FTS5 search).

Cost first by design: text only analysis, free transcription, no multimodal model. The original build sent every video to a multimodal model and paid per minute of transcription. Those paths are gone.

Built by an AI agent team

Snag is developed by a small team of AI agents orchestrated through Hermes and a Kanban board. Work lands as pull requests: a Builder agent opens a feature branch and a PR, a Verify agent runs QA, and changes merge on approval. Every feature commit in this repo is the output of that loop.

Run it

cp .env.example .env            # fill in your keys
python3 test_pipeline.py "https://www.tiktok.com/t/XXXX"
python3 bot.py

Tests:

pytest -q                        # 157 tests

See .env.example for the required keys. Transcription needs faster-whisper and ffmpeg. TikTok download fallback needs yt-dlp.

Find Snag

• Telegram: @snagcapturebot
• X: @snagcapture