PeakCut: turn long videos and live streams into captioned vertical clips (free Opus Clip alternative)
At the end of the article you'll find the install command, a ready-to-paste prompt and a .md to download.
PeakCut is an open source app that turns long videos and live streams into 9:16 vertical clips with word-by-word captions, an on-screen hook and a ready-to-post title. Paste a YouTube, Twitch or Kick link (or upload a file), transcription runs on your own machine, and an AI model of your choice decides which moments deserve a clip. It does the same job as paid services like Opus Clip, but it runs on your PC or server, it is MIT licensed, and you only pay for what your AI key uses (or nothing, with a local model).
We installed it and tested it with two videos, one in English and one in Spanish. This guide covers how it works inside, how to install it, what it really costs, and the one setting you need for non-English captions, because the default gets them wrong.
What PeakCut is and who makes it
- Repository: github.com/haziqahmedqasmi/PeakCut. There are no numbered releases: we reviewed commit
1bef2b4from November 17, 2026. The app itself shows a notice in the dashboard when a newer one is out. - License: MIT. The caption fonts (Anton, Archivo Black, Poppins) are under the SIL Open Font License. You can use it for client work at no cost.
- Stack: a Python server on FastAPI,
ffmpegfor cutting and rendering,yt-dlpfor downloads,sherpa-onnxwith the Parakeet model for speech to text, and a plain HTML, CSS and JavaScript interface with no build step. SQLite on your PC, PostgreSQL on a server. - What you get: clips from any video, real-time live stream clipping, per-creator workspaces, a posting queue for YouTube, Instagram and TikTok with spacing and daily caps, and an AI usage page with requests and tokens per video.
How it picks the moments to clip
Each video goes through five steps, worth knowing so you know what to tune:
- Download with
yt-dlp(known video sites only), or read the file you uploaded. - Transcribe locally with Parakeet 0.6b. Audio never leaves your machine; the model is about 480 MB and downloads once.
- Ask the AI which moments would hold a stranger's attention. The prompt lives in
studio/brain.py: start on the setup line, end right after the payoff, skip intros and sponsor reads, and return fewer clips when there are no good ones. The README estimates about four requests per hour of video. - Snap the cuts so each clip starts at the beginning of a line and ends where a sentence ends.
- Render each clip at 1080×1920 with
ffmpeg: captions with the active word highlighted, the hook on top and the layout you chose.
It also works with no AI key: it picks the loudest stretches that contain speech. The judgement is worse, but it lets you try it for free.
For live streams it records into a rolling buffer and clips while the stream is still running, combining loudness, chat activity and the transcript. Chat is counted by distinct people rather than messages, so one account spamming emotes does not look like an excited crowd. Gaming streams get a gaming mode (GAMING.md): it recognises 23 games from the title, adds their hashtags, and finds the webcam by looking for a straight border that stays put between frames, to build the "Gameplay + facecam" layout.
Install PeakCut on your PC
You need Python 3.10 or newer and ffmpeg:
# Windows
winget install Gyan.FFmpeg
# macOS
brew install ffmpeg
# Linux (Debian/Ubuntu)
sudo apt install ffmpegClone the repo into a short, permanent folder. On Windows this matters: if the full path goes over 260 characters, the speech library will not load.
git clone https://github.com/haziqahmedqasmi/PeakCut C:\peakcut
cd C:\peakcut
pip install -r requirements.txt
python -m studio.serverOn macOS or Linux use ~/peakcut. Open http://127.0.0.1:8765: there is no sign-in, it opens straight into the dashboard with an 8-step tour. If you ask Claude Code or another agent to install it, the repo's AGENTS.md tells it to ask you where to put it first, so it does not end up in a temporary folder that gets wiped.
To try it without touching real data, the repo suggests running a copy on another port and folder:
PEAKCUT_DATA=/tmp/peakcut-test PEAKCUT_PORT=8766 python -m studio.serverChoosing the AI and what it really costs
In Settings → AI pick the service, paste the key, press Find models to see the provider's real list, then Save and test. The key is stored in the data folder on your machine and never shown again.
| Option | What you pay | When to choose it |
|---|---|---|
| No key | Nothing | Trying the app. Picks by loudness, no good titles or hooks. |
| Ollama (local model) | Nothing but your hardware | You have a decent GPU and want nothing to leave your machine. |
| Gemini or Groq | Their free tier, with usage limits | The cheapest option with a cloud model. |
| OpenAI, Claude, Grok, OpenRouter | Whatever you use on their API | You already have an account and credit with one of them. |
| Custom | Depends on the provider | Any API that speaks the OpenAI chat format. |
Two practical details: the README recommends a fast, inexpensive model, and you can list several, comma-separated, in PEAKCUT_AI_MODELS so it falls back to the next one if the first fails or rate-limits you. Transcription does not use the AI and costs nothing. Compared with a subscription clipping service, PeakCut's cost comes down to your electricity and, with a cloud provider, a few requests per hour of video.
Real test: an English video and a Spanish live stream
We built two test videos with synthetic voice: a 1:51 talk in English and a 1:21 "stream" in Spanish. In a Linux container with no GPU, the English one was transcribed, cut and rendered in 31 seconds, according to the dashboard. With no AI key, PeakCut picked a 57-second stretch by loudness and gave it the generic hook "WAIT FOR IT".
The Spanish video exposed the important issue. The default model is parakeet-tdt-0.6b-v2, which only understands English: it wrote misspelled Spanish and, halfway through, drifted into English with lines like "The problem is the vilidad". Parakeet v3 is multilingual and covers 25 European languages, and PeakCut lets you switch models with one variable:
PEAKCUT_ASR_MODEL=sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8With that line in .env, it downloaded the new model on startup (another ~480 MB) and the transcript came out clean, with accents and punctuation.
Two more things we read in the code that matter for non-English content:
- With the AI on, the hook, title, description and hashtags are requested in English (except Hindi and Urdu). It is written in the prompt in
studio/brain.py; since it is MIT, you can change that line to your language. - If the video comes from YouTube and YouTube reports another language, the clip's captions are translated into English. Uploaded files, Twitch and Kick keep the spoken words.
Posting to YouTube, Instagram and TikTok
In Settings → Connected apps each user presses Connect, signs in on the platform and comes back. For that, the server needs one registered app per network, and that is the slow part, because it does not depend on PeakCut:
- YouTube: a Google Cloud OAuth client with YouTube Data API v3. Until Google verifies the app, uploads stay private.
- Instagram: a Business app in Meta for Developers. Works with Creator and Business accounts; until Meta reviews the app only tester accounts can connect, and Instagram fetches each clip from a signed link on your site, so you need a public https address.
- TikTok: an app with Login Kit and the Content Posting API. Until TikTok audits the app, clips are posted as private.
If you clip for several creators and would rather not fight Meta, Google and TikTok reviews yourself, an agency like Safe Build Labs can set up the server and the posting apps for you. Once the apps are approved, the queue posts with spacing between clips and a daily cap, and the Pause all posting button stops all of it.
Running it on a server with Docker
The repo ships a docker-compose.yml with three containers: the app, PostgreSQL and Caddy for automatic HTTPS. There are Traefik and Nginx variants in deploy/, so it fits a VPS running Coolify.
cp .env.example .env # AI, PEAKCUT_DOMAIN, POSTGRES_PASSWORD
docker compose up -d --build
docker compose exec app python -m studio.admin grant you@example.comDEPLOY.md asks for 8 cores, 16 GB of RAM and 200+ GB of disk if several people will use it, because clipping is CPU work. Transcription runs at roughly 10x real time on 4 cores; if that is not enough, deploy/gpu/ sets up a separate NVIDIA machine just for speech (PEAKCUT_ASR_URL and PEAKCUT_ASR_TOKEN).
The security rule you cannot skip: without accounts, PeakCut only listens on 127.0.0.1. On any server reachable from the internet set PEAKCUT_ACCOUNTS=1 (Docker already does). Admin roles can only be changed with the studio.admin command on the machine, never from the web. For the same deployment pattern with another app, see how we installed Paymenter on a VPS with Coolify.
Common errors
- "DLL load failed ... The filename or extension is too long" on Windows: the folder is nested too deep. Move it to
C:\peakcut(copy.env,storage.jsonanddataif you already have clips). - Made-up English captions on a non-English video: set
PEAKCUT_ASR_MODELto the v3 model. - "YouTube isn't letting PeakCut download videos right now": YouTube sometimes blocks downloads from servers. Upload the file or use a Twitch or Kick link; PeakCut retries YouTube a few minutes later.
- Every clip says "WAIT FOR IT": no AI is set up, so it is picking by loudness. Add a key in Settings → AI.
- Disk full: PeakCut stops taking new videos until there is room. Settings → Storage lets you move the data folder to another drive.
FAQ
Is PeakCut free? Yes. The code is MIT and transcription is local. You only pay for a cloud AI if you pick one; with Ollama or no key you pay nothing.
Does it work in other languages? Yes, if you switch the speech model to Parakeet v3. The AI's hooks and titles come out in English unless you edit the prompt.
Do I need a GPU? No. It runs on CPU; an NVIDIA GPU only speeds up transcription on busy servers.
Can I clip other people's videos? Technically yes, but the README is clear: clip what you made or have permission to use. Rights and each platform's terms are on you.
How is it different from Opus Clip? The output is similar. The difference is that PeakCut is yours: it runs on your machine, you choose the AI model, there is no minutes cap, and you can read and change how it picks cuts.
Verdict
PeakCut is the most complete open source option we have tried for getting Shorts, Reels and TikToks out of long videos and live streams: local transcription, any AI you like, gaming mode and automatic posting. For non-English content, set PEAKCUT_ASR_MODEL to the v3 model from day one and decide whether you want English titles or an edited prompt. Next step: install it on your PC, try one of your own videos with no key, and once you are convinced, connect Gemini or a local model. If you also want to generate videos from scratch with agents, see how to make AI videos with Claude Code or Codex and OpenMontage.
// resource · use it now
PeakCut ready to use on your PC or server
Install it by hand with the commands, or paste the prompt for your case into Claude Code (or another agent) to get it running, including non-English captions.
Repository license: MIT (caption fonts: SIL OFL)
1. Install
winget install Gyan.FFmpeg
git clone https://github.com/haziqahmedqasmi/PeakCut C:\peakcut
cd C:\peakcut
pip install -r requirements.txt
python -m studio.servergit clone https://github.com/haziqahmedqasmi/PeakCut ~/peakcut
cd ~/peakcut
pip install -r requirements.txt
python -m studio.serverPEAKCUT_ASR_MODEL=sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int82. Pick your case
I want to install PeakCut (https://github.com/haziqahmedqasmi/PeakCut) on this PC to make vertical clips from my videos.
1. Before cloning, ask me for the folder. Suggest C:\peakcut on Windows or ~/peakcut on macOS/Linux, never a temporary folder or your own workspace.
2. Check that I have Python 3.10+ and ffmpeg; if ffmpeg is missing, give me the install command for my system and wait for my OK.
3. Clone the repo, create a virtual environment and install requirements.txt.
4. If my videos are not in English, create a .env from .env.example with PEAKCUT_ASR_MODEL=sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8.
5. Start python -m studio.server, wait for "Speech model ready" and give me the URL.
6. Do not put any AI key in files: I will add it myself in Settings → AI.3. How to use it
- Open the URL the agent gives you (
http://127.0.0.1:8765) and follow the 8-step tour. - Upload a short video of your own with no AI key to check the captions.
- Add your key in **Settings → AI** (Gemini or Groq to start cheap, Ollama to keep everything local) and press **Save and test**.
- For hooks and titles in your language, ask the agent to change the line "Write hook, title, description, hashtags and highlight_words in English" in
studio/brain.py.
I want to deploy PeakCut (https://github.com/haziqahmedqasmi/PeakCut) on my VPS with Docker for several creators.
1. Read the repo's README.md, DEPLOY.md and AGENTS.md before touching anything; their rules win.
2. Prepare .env from .env.example: PEAKCUT_DOMAIN, POSTGRES_PASSWORD, PEAKCUT_ACCOUNTS=1, and PEAKCUT_ASR_MODEL=sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8 if the content is not in English. Leave keys and secrets empty: I will add them in the dashboard or the server's variables, never in the repo or the chat.
3. If the server already runs Coolify, Traefik or Nginx, use the matching variant in deploy/ instead of Caddy and tell me why.
4. Check the server meets DEPLOY.md's requirements (cores, RAM, disk) and warn me if it falls short.
5. Run docker compose up -d --build, wait until the app answers over https, and give me the studio.admin grant command to make myself admin.
6. Finish with a list of what is left: the YouTube, Instagram and TikTok OAuth apps and which review each one needs.3. How to use it
- Open your domain, create your account and run the
studio.admin grantcommand the agent gave you on the server. - Add the AI key in **Settings → AI** and create one workspace per creator.
- Register the posting apps from **Settings → Connected apps → Set up**: the redirect address is shown with a Copy button.
- Until Google, Meta and TikTok review your apps, post privately or with tester accounts and review clips before turning on auto-posting.
Help us grow
codigoagente is free, with no paywalls. If this article helped you, any of these keeps the blog going.
- Share it with someone who'll find it usefulWhatsAppLinkedInXFacebook
- Meet our sponsorSafe Build Labs
More on Open source tools

- Free alternatives to paid apps
- Design, image and video

- Claude Code
- Free alternatives to paid apps

- Claude Code
- Free alternatives to paid apps