One long video in. Scroll-stopping shorts out, in English, Spanish and Portuguese.
One long video in. Finished English, Spanish and Portuguese shorts out, with captions, thumbnails and copy ready for review. Every step is a free open-source tool except the two AI steps.
yt-dlp pulls the original video. The source file is never edited or overwritten.
yt-dlpfaster-whisper writes a word-by-word transcript with exact timings (JSON + SRT).
faster-whisperAn AI reads the transcript and picks the strongest 20–45s moments. Timecodes are checked against real words; invented ones get rejected.
LLM + checkerEach moment is cut to 9:16 from the original, with bold word-timed captions (2 lines, 42 characters max) and a hook title.
ffmpeg + libassSpanish (LatAm) and Brazilian Portuguese with the same cue timing, a locked glossary of your phrases, and an automatic check for line length and glossary terms.
LLM + glossaryEvery clip in every language gets an MP4, SRT, thumbnail, title, caption, hashtags and alt text. A frame from every caption is checked. Nothing is posted until you say yes.
ffmpeg + checksSource: Create Content Quickly (for Busy Business Owners). Tap a language to switch the video, captions and copy. These are drafts for your review. Nothing has been posted anywhere.
Method: `ecm.py verify` pulled one frame per cue (cue midpoint) from each finished mp4 and laid the caption band
next to the expected .srt text (contact sheets: verify/<clip>/<lang>/sheet.jpg, combined: verify/<clip>-all.jpg).
I read every row of all 9 sheets.
Result: 165 / 165 cue frames show exactly the .srt text, same line breaks, no missing, extra or stale cues.
- 01-set-a-deadline: en 13/13, es 13/13, pt-BR 13/13
- 02-one-weekend-course-and-book: en 27/27, es 27/27, pt-BR 27/27
- 03-dare-to-suck: en 15/15, es 15/15, pt-BR 15/15
Accents render correctly (é, ñ, ã, ç, ¿, ¡). Source's own English captions/lower-thirds are trimmed out of frame
(runs/2k01Q2Vvr2A/overrides.json), so ES/PT shorts show no English caption text.
Note: in 03 cue 5 the source's own big orange "RIP IT" graphic is partly visible at the edge of the frame (it is part of the video picture, not our captions).
Short answer: yes. Everything is permissive except FFmpeg/x264 (GPL in this build). We can call FFmpeg freely and own the videos it makes. If we ship the FFmpeg binary inside a paid bundle, we also have to offer its source code. Checked Nov 18 2026: license field from the GitHub API for each repo, plus the FFmpeg LICENSE.md, x264 COPYING and Pillow LICENSE files and the Hugging Face model card.
| Tool | License | What it means for bundling / reselling |
|---|---|---|
| yt-dlp 1. Fetch (download) | Unlicense (public domain) | Free to bundle, rebrand and resell. No notice required. Only download videos you own or have rights to (YouTube terms). source |
| faster-whisper 2. Transcribe | MIT | Free to bundle and resell. Keep the MIT copyright notice in the bundle. source |
| CTranslate2 2. Transcribe (engine) | MIT | Free to bundle and resell with the notice. source |
| Whisper model weights (Systran/faster-whisper-medium, from OpenAI Whisper) 2. Transcribe (model) | MIT | Free to ship and use commercially with the notice. source |
| FFmpeg (this box's build) 4/7. Cut + burn captions | LGPL-2.1+ by default; GPL-2.0+ when built with --enable-gpl. This box's build has --enable-gpl and libx264, so it is GPL. | Calling ffmpeg as a separate program is fine, and the videos it makes are yours. If we ship the ffmpeg binary inside a paid bundle we must also offer its source code (GPL). Cleanest: the installer has the customer install ffmpeg, or we ship an LGPL build. The pipeline avoids GPL-only filters (uses gblur, not boxblur). source |
| x264 (H.264 encoder inside ffmpeg) 7. Burn / encode | GPL-2.0 | Same GPL rule as above if we ship it. x264 also sells a commercial license (UNVERIFIED: not checked today). H.264 patent licensing is a separate question (UNVERIFIED). source |
| libass 7. Caption renderer | ISC | Permissive: bundle and resell with the notice. source |
| Pillow Text fitting, verify sheets | MIT-CMU / HPND-style permissive (GitHub API says NOASSERTION; read from LICENSE file) | Permissive: bundle and resell with the notice. source |
| PyAV Installed with faster-whisper | BSD-3-Clause | Permissive. Note its wheels bundle FFmpeg libraries (LGPL). source |
| NumPy Audio buffer | BSD-3-Clause (UNVERIFIED today) UNVERIFIED | Expected permissive; not checked today. source |
| Montserrat font (caption font; static ExtraBold instance renamed 'Montserrat ECM') Captions | SIL OFL 1.1 | Can ship inside paid software. Can't sell the font by itself. Our modified copy must stay under OFL. source |
| Anton font Hook titles | SIL OFL 1.1 | Same as Montserrat. source |
| Wrangler (Cloudflare deploy CLI) Deploy only | Apache-2.0 (workers-sdk repo) | Not shipped in the bundle; used to publish this page. source |
| LLM for pick + translate (Grok Bot today; Claude or any model via ECM_LLM_CMD) 3. Pick, 5. Translate + review | Not open source. Paid per use. | The one step with a running cost. Customers bring their own model key, or Epic covers usage. |
| Epic Clip Machine code (ecm.py, burn.sh, site) The machine | Epic's own code. No license file yet: Trevor's call (private / proprietary bundle). | Ours to brand and sell. |
This is a plain-English summary, not legal advice.
On the box, one command. It stops and asks for the two AI answers (moment picks, translations) unless an AI command is set with ECM_LLM_CMD; re-running picks up where it left off.
cd /workspace/epic-clip-machine python3 ecm.py run "https://www.youtube.com/watch?v=VIDEO_ID" # optional: let a CLI model answer the AI steps automatically ECM_LLM_CMD='claude -p' python3 ecm.py run "https://www.youtube.com/watch?v=VIDEO_ID" # rebuild + redeploy this page python3 site/build_site.py && ./deploy.sh