Est. by dead-angel · hosted, or run it yourself
ARCH‑ANGEL downloads it, listens to it, decides which thirty seconds actually matter — then cuts, captions, reframes and scores them. No per-clip metering, no watermark, no queue to buy your way out of.
…or run the whole thing on your own box. same software.
Output landed in output/podcast-47/ with a
.json sidecar per clip. Nothing left the building.
The going rate is a monthly fee plus a per-minute charge, and the deal is always the same: upload your raw video to a stranger's bucket and hope the terms don't change. That's a bad trade for anyone with a back catalogue and a GPU sitting idle.
So ARCH‑ANGEL does it locally. yt-dlp fetches, faster-whisper transcribes, and a picker you choose — plain heuristics, audio energy, a local Ollama model, or a cloud LLM if you feel like it — argues about which moments are worth keeping. ffmpeg does the cutting. Clip as much as you like; nothing counts down.
§ 02 — the assembly line
You can stop after any of them. --dry-run gets you the moment
list without burning a single frame of render time.
§ 03 — what's in the box
Finished a clip and want it different? Open it. A real waveform timeline with draggable trim handles, live grade preview, caption and watermark overlays that update as you drag — then re-render. Nothing is destructive; the original stays put.
Point it at a creator's channel. Every new upload gets clipped automatically using that channel's own profile — its picker, its caption style, its music, its clip length — and can auto-publish as a Short or a regular video at whatever privacy you chose in advance.
Connect as many YouTube channels as you run. Switch the publish target in one click. Each account's tokens stay private to it.
Tick two or more pickers and the moments they agree on rank highest. Consensus beats a single opinionated model.
Upload your own music beds and SFX stings, reorder them by dragging, and they show up everywhere a track can be picked.
Real whisper timings, five animation styles, any font, your colours.
A super admin holds the API keys. Everyone else just makes clips.
Render, upload and server failures land in one readable place.
Clips per day, score spread, library growth — from your own output.
§ 04 — the small print
| Runtime | Python 3.10+, stdlib HTTP server (no Flask) |
|---|---|
| Fetch | yt-dlp |
| Transcribe | faster-whisper · CUDA when present |
| Render | ffmpeg (a Windows build is bundled) |
| Pickers | heuristic · energy · Ollama · 7 cloud LLMs |
| Captions | ASS subtitles, 5 animation modes |
| Publish | YouTube Data API v3, browser OAuth + PKCE |
| Auth | PBKDF2-SHA256, HttpOnly sessions, HTTPS + HSTS |
| Deploy | Docker or systemd · self-host in minutes |
| Lock-in | None. Take the software and go. |
§ 05 — two ways in
Sign in and paste a link. Accounts are made by an administrator, so open the Studio if you have one — otherwise ask for access.
# install once
pip install -r requirements.txt
# clip a URL, five shorts
python main.py "https://youtube.com/watch?v=..." --clips 5
# or run the whole workshop locally
python dashboard.py # -> http://localhost:8000
# going live on your own domain? one file, one command
cp .env.example .env && docker compose up -d --buildEvery flag, the YouTube setup and the deployment guide live in the manual.
§ 06 — asked often
tiny or base whisper model and be patient. The
Studio tells you which engine it detected.go on then. →