Turn any YouTube video into viral short-form clips — fully automatic.
AI finds the highest-engagement moments, cuts the video, crops to 9:16 while tracking the active speaker, and burns word-by-word karaoke captions. Clips can be downloaded or posted straight to YouTube Shorts and TikTok — manually or fully hands-off via channel monitoring.
Live at clipforging.com.
Clip generation
- AI clip extraction — OpenRouter model scores every transcript segment for virality (hooks, tension, confessions, numbers, story arcs) and picks the best moments; automatically falls back to Groq Llama if OpenRouter is unset or unavailable
- Word-by-word karaoke captions — three styles (Bold Bottom, Center Pop, Minimal), custom font size, custom highlight color
- Caption translation — 15 languages (Arabic, Chinese, Dutch, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Turkish, Ukrainian)
- Smart 9:16 reframe — YOLO + audio-gated active-speaker tracking pans to whoever is talking, with dead-zone smoothing to avoid jitter; no manual cropping
- Hardcoded-caption detection — detects and crops out burned-in subtitle bars so they don't end up doubled in the clip
- Scene-aware boundaries — snaps clip cuts to scene changes so clips don't start/end mid-shot
- Silence trimming — strips dead air from the ends of clips
- Background music — mix a track (from a YouTube URL) underneath clips
- Custom style prompt — steer the AI toward a specific clip style or topic
- Auto thumbnails — a thumbnail is generated for every clip
Publishing
- YouTube auto-upload — OAuth-linked; connect multiple channels and upload clips straight to any of them
- TikTok cross-posting — Login Kit + Content Posting API (Direct Post); connect multiple accounts, pick privacy live from creator settings
- Per-clip analytics — views, likes, and comments from the YouTube Data API, auto-refreshed every 6 hours
Automation
- Watchlist / channel monitoring — add any YouTube channel; ClipForge detects new uploads and clips them automatically (Pro)
- Channel Digest — backfill a channel's historical videos on a daily schedule with a look-back window (30/60/90 days, 6mo, 1yr) and a per-day cap; auto-stops when caught up
- Auto-upload — watchlist and digest clips can publish themselves to a chosen YouTube channel and/or TikTok account
- Email notifications — Resend email when your clips are ready
Platform
- Job archive — every job stored with full settings; one-click retry carries all original parameters; cancel a running job or delete one outright
- Cloudflare R2 storage — clips served via short-lived presigned URLs; auto-deleted after 7 days
- Anti-abuse — disposable/temp-email blocking and Gmail alias normalization at signup
- Free / Pro tiers — Free: 10 clips/month, 3 per job. Pro: unlimited (coming soon)
┌─────────────────────────────────────┐
│ 🎬 YouTube URL │
└──────────────────┬──────────────────┘
│
╔════════▼════════╗
║ 📥 DOWNLOAD ║ yt-dlp — best quality video + audio
╚════════╤════════╝
│
╔════════▼════════╗
║ 🔀 MERGE ║ FFmpeg — combines streams into mp4
╚════════╤════════╝
│
╔════════▼════════╗
║ 🎙 TRANSCRIBE ║ Groq Whisper (whisper-large-v3)
║ ║ word-level timestamps · hallucination filter
║ ║ auto-chunks files > 20 MB
╚════════╤════════╝
│
╔════════▼════════╗
║ 🧠 ANALYZE ║ OpenRouter (Groq Llama fallback)
║ ║ multi-signal virality scoring
║ ║ optional caption translation
╚════════╤════════╝
│
╔════════▼════════╗
║ ✂️ CLIP ║ FFmpeg — cuts on scene boundaries
║ ║ YOLO active-speaker 9:16 crop
║ ║ caption-bar crop · silence trim
║ ║ burns ASS captions · bg music · thumb
╚════════╤════════╝
│
┌──────────────────▼──────────────────┐
│ ☁️ Stored in Cloudflare R2 │
│ 🚀 Download · YouTube · TikTok │
└─────────────────────────────────────┘
| Layer | Technology |
|---|---|
| Frontend | React + Vite, React Router v6 |
| Backend | FastAPI (Python) |
| Auth & Database | Supabase (Postgres + Auth) |
| Clip Storage | Cloudflare R2 (S3-compatible) |
| Transcription | Groq Whisper (whisper-large-v3) |
| AI Analysis | OpenRouter (primary) → Groq Llama (fallback) |
| Video processing | FFmpeg, yt-dlp, OpenCV + Ultralytics YOLO |
| Publishing | YouTube Data API, TikTok Content Posting API |
| Resend |
- Python 3.10+, Node.js 18+,
ffmpeg,yt-dlp - A Groq API key (free tier works) — transcription + analysis fallback
- A Supabase project
- A Cloudflare R2 bucket
- Optional: OpenRouter (primary analysis model), YouTube OAuth + Data API, a TikTok developer app, and a Resend key for emails
chmod +x setup.sh start.sh
bash setup.shCopy the two templates and fill in your values:
cp .env.example .env # backend secrets
cp frontend/.env.example frontend/.env.local # Vite (browser) configfrontend/.env.localholds only the twoVITE_Supabase values (these are exposed to the browser — never put a secret here)..envholds everything else. Only the four Required values plus the R2 block are needed to boot; every other section (OpenRouter, YouTube, TikTok, Resend, cookies) is optional and unlocks the matching feature when set. See.env.examplefor the full annotated list.
Where
.envis loaded from: the backend reads.envfrom the clipforge root (one level above theclipper/folder) in local dev, or/home/ubuntu/.envon the server.COOKIES_FROM_BROWSER: when set, yt-dlp reads live cookies directly from the browser profile — they never expire. If not set, it falls back to a
cookies.txtfile at the clipforge root.
bash start.shOpen http://localhost:8000, create an account, and start forging clips.
| Stage | What happens |
|---|---|
| DOWNLOAD | yt-dlp fetches best quality video + audio |
| MERGE | FFmpeg combines streams into a single mp4 |
| TRANSCRIBE | Groq Whisper generates word-level timestamps |
| ANALYZE | OpenRouter (Groq Llama fallback) scores moments for virality; optionally translates captions |
| CLIP | FFmpeg cuts on scene boundaries, crops to 9:16 with active-speaker tracking, removes caption bars, trims silence, burns karaoke captions, mixes background music, and generates a thumbnail |
yt-dlp fails: pip install -U yt-dlp
FFmpeg not found: sudo apt install ffmpeg (Linux) or winget install Gyan.FFmpeg (Windows)
Groq rate limits: Add GROQ_API_KEY_2 … GROQ_API_KEY_5 to spread load across keys automatically
Supabase 400 on insert: Make sure all columns exist in the database — PostgREST silently drops unknown fields
R2 clips not loading: Check R2_ACCOUNT_ID, R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, and R2_BUCKET_NAME are all set
Analysis using the wrong model: OpenRouter is primary only when OPENROUTER_API_KEY is set; otherwise it silently uses Groq Llama. Check the key and OPENROUTER_MODEL.
YouTube "Sign in to confirm you're not a bot": set COOKIES_FROM_BROWSER (or provide cookies.txt), and optionally run the PO-token sidecar and set POTTOKEN_URL.
TikTok posts only to private: while the app is unaudited, TikTok forces SELF_ONLY. After production approval, set TIKTOK_PRIVACY_LEVEL=PUBLIC_TO_EVERYONE.
Digest jobs fail with "Server disconnected": fixed — concurrent pipelines no longer share a corrupted Supabase connection (the DB layer retries on a fresh one).