NeonClipping turns long streams into vertical Shorts on your own PC, without uploading the source file to a vendor cloud. The GitHub repository by Ajay K Meena gained more than 1,700 stars within days of its November 19, 2026 creation, riding the same wave as other open clipper repos but emphasizing audio-first partial downloads, gaming layouts, and local Ollama titling.
The README's headline benchmark: clip an 85-minute YouTube stream in 97 seconds on an RTX 2060, with zero API cost in Fast mode when YouTube captions exist.
Quick facts (per the README)
| Question | Answer |
|---|---|
| Input | YouTube, Twitch, Kick URLs or local MP4 |
| Output | 9:16 MP4 with burned-in captions, title, hook |
| Cost | Free; local compute only |
| Default mode | Fast mode for videos over 10 minutes |
| LLM | Local Ollama for moment scoring and copy |
| Privacy | Server binds 127.0.0.1; README says nothing is uploaded |
| Install | Windows one-click .bat, install.sh on Mac/Linux |
| Stars (Nov 22) | ~1,733 on GitHub trending feed |
Why another open-source clipper exists
Short-form clipping is crowded. We already covered AutoClip, which ships desktop, Docker, CLI, and MCP surfaces and bills cloud models per README measurements. NeonClipping chases a different user: streamers and editors who want facecam + gameplay layouts, live stream clipping, and very long VODs without pulling the entire video file first.
The README comparison table positions NeonClipping against OpusClip (cloud subscription) and Docker self-hosted clippers (SupoClip, OpenShorts, Reelsaw). Its claimed advantages: no Docker, multi-platform live links, 1080p default with 4K option, 11 layout presets, and chat-spike based live clipping with a phone remote.
None of that removes legal risk. The project is explicit that you should star the repo if it saves time; it does not grant rights to re-upload third-party gameplay.
How Fast mode stays fast
NeonClipping's Fast mode copies the strategy the README attributes to the fastest clippers: never analyze the whole video file up front.
- Download audio only (README cites 79 MB for 85 minutes, ~10 seconds).
- Use YouTube captions when present—word-timed, no speech model. Fall back to local Whisper on other sources.
- Score the audio track for loudness jumps vs a 2-minute baseline, raised voice energy in 2–4 kHz, and quiet-to-loud transitions.
- Send top moments to Ollama to score emotion, write a title and hook, snapping cuts to sentence boundaries within min/max length sliders.
- Download only chosen segments of video, align trims to the audio track, render 1080×1920 with captions.
Deep mode downloads the full video, adds OpenCV Farneback optical flow for motion peaks, shortlists 4× more candidates, and gives the LLM 15 seconds of context on each side.
Measured times on one machine (README):
| Source | Fast mode time |
|---|---|
| 85-minute YouTube (captions present) | 97 s |
| 3-hour Twitch VOD (local ASR) | 136 s |
| 90-second YouTube trailer | 21 s |
| 90-second local MP4 | 14 s |
Highlight scoring and on-device learning
Candidates combine loudness jumps, shout band energy, keyword hits (English, Hindi, Hinglish examples in README), optional motion and flash cues in Deep mode, and the Ollama score.
Ratings on generated clips feed a local data/feedback.jsonl file. After enough diverse five-star feedback, the README says local ranking weights can adapt—preference learning, not fine-tuning the foundation model. Nothing uploads.
Caption tooling supports GTA Neon, Impact, Clean, Hormozi, MrBeast styles, 52 languages, Hinglish romanization defaults, and an opening heading block with optional Improve opening heading using transcript context.
Gaming and live stream features
For creators, the layout table matters more than benchmark points:
- 11 native layouts (camera above/below gameplay, blurred fill, circle facecam, side-by-side, gameplay-only, etc.).
- Facecam finder plus manual box drawing in the editor.
- Live stream clipping using chat-spike signals and a phone remote, rare in self-hosted clippers.
- Re-edit without re-downloading: change layouts/caption styles on saved clips if source files remain in
data/.
Optional Gemma4 visual review samples three frames per candidate when installed; default off to save CPU.
Install paths and agent setup
Windows: download NeonClipping-Setup.bat from the project site or Releases. The script checks Python and FFmpeg, installs into %LOCALAPPDATA%\NeonClipping, and opens the UI.
Mac/Linux: install.sh exists; README notes it is not yet tested on a real Mac.
Agents: copy docs/NEONCLIPPING-AGENT-SETUP-PROMPT.txt into Claude Code, Codex CLI, or Cursor. That fits the pattern from our Viggle MCP video agent post: give agents a bounded install prompt, then supervise first run.
Phone grid of short video tiles, illustrating NeonClipping local YouTube Shorts output
NeonClipping vs AutoClip for builders
| Axis | NeonClipping | AutoClip (our Nov 19 coverage) |
|---|---|---|
| Agent interface | Setup prompt text | First-class MCP server |
| Partial download | Core Fast mode design | Desktop/local render focus |
| Live streams | Chat-spike + phone remote | Not highlighted |
| Gaming layouts | 11 presets | General interview/podcast |
| Cloud spend | Avoids by design | README lists small yuan/model run costs |
Pick NeonClipping when Twitch or Kick facecam workflows and multi-hour VODs dominate. Pick AutoClip when you want MCP in an agent cron clipping podcast folders.
Risks and limits to verify
- Platform ToS: Downloading streams may violate YouTube or Twitch terms depending on how you use outputs. The tool does not legal-check your use case.
- Caption quality: Fast mode leans on YouTube captions when present; bad auto-captions become bad hooks.
- Ollama dependency: Titles and hooks require a working local model; budget disk for weights.
- Pre-release ergonomics: Mac installer lacks field testing per README.
- Security: Localhost binding reduces exposure but the app still fetches remote URLs you supply—do not run unsupervised on sensitive networks.
Storyboard frames, showing NeonClipping re-edit without redownloading source
What to do next
- Run Fast mode on your own VOD before client footage.
- Pin Ollama model versions used for hooks.
- Compare output retention against a manual editor baseline—automated loudness peaks are not comedy timing.
- If you need agent pipelines at scale, evaluate MCP-first tools like AutoClip alongside NeonClipping's prompt installer.
Teams monetizing Shorts should still read platform music and gameplay policies; local clipping does not exempt you.
A practical first-week plan helps you judge the tool fairly. On day one, clip a single stream you own in Fast mode and note the wall-clock time against the README's 97-second figure for your own GPU. On day two, run the same source in Deep mode and compare which moments each mode picked. Deep mode downloads the full video and adds optical-flow motion peaks, so it should surface action beats that audio loudness alone misses. On day three, rate every clip so the local feedback file has enough diverse scores to start adapting rankings. By the end of the week you will know whether audio-first scoring matches your content style, or whether you need Deep mode as your default.
Honest limits from the README
NeonClipping is explicit about edge cases builders should expect in production.
Language and captions. Fast mode prefers YouTube captions; when they are missing, local Whisper large-v3 on Hindi can be slow on small GPUs. Imported SRT/ASS/WebVTT timings are estimated inside cues, not guaranteed frame-perfect.
Visual review uncertainty. Optional Gemma image review "does not verify a complete game event," per README. Do not treat three-frame summaries as ground truth for esports highlights.
Feedback learning threshold. Local weight adaptation needs at least eight rated moments with diverse scores; until then ranking stays on default heuristics.
Hardware exports. Validation notes cite seven layout/style re-exports on one 796.9-second source in 119.58 seconds—helpful as a sanity check, not a service-level agreement.
Regression tests. The repo ships python -m unittest discover -s tests -v for portable checks before you fork the UI.
Where NeonClipping sits in the agent video stack
Cloud clipper APIs optimize for non-technical marketers. Agent-native repos optimize for scriptable batch windows and MCP. NeonClipping adds a third lane: solo creators who want DaVinci-grade layout control without shipping footage to a SaaS transcode farm.
That is why the README links Meena's background as a director, DOP, and Resolve educator in Jaipur: the UI targets editors who already think in timelines, not only Python agents. The included English and Hindi PDF guide signals the audience.
If you are building a multi-tenant clipping service, NeonClipping's localhost architecture is a prototype, not a host. Run it as a desktop helper, or fork the scoring logic into your own worker pool with explicit consent flows.
For rights-sensitive enterprises, pair local clipping with our token fraud and trial-abuse mindset: free local tools still cost support time when users paste pirated feeds.
FAQ addendum: operational checklist
Before you publish NeonClipping output commercially:
- Confirm broadcast rights for game footage and music beds burned into captions.
- Log which Ollama model wrote titles if you need to reproduce copy for a dispute.
- Keep source files in
data/until re-edits finish; the README warns exports fail if originals are deleted. - Schedule disk cleanup; partial downloads still accumulate for saved projects.
Trending context on GitHub
The repository shows 1,733 stars in the November 22 candidate feed, with 54 commits by then. Trending alone does not prove retention; it signals that local, no-upload clipping still beats subscription SaaS for a large creator slice. Watch whether the project adds MCP or HTTP APIs—today it is UI-first with an agent install prompt, similar to early AutoClip before MCP landed.
If you maintain a clipping fork, read the upstream MIT License before you commercialize hosted rendering.
