Turn long videos and live streams into captioned vertical clips, ready to post. Free and open source: run it on your own computer or server, with your own AI key.
Paste a YouTube, Twitch or Kick link (or upload a file) and PeakCut finds the moments worth clipping, cuts them to 9:16, adds word-by-word captions and a hook, and can post them to your connected accounts. It can also record a creator's live stream and clip the highlights while the stream is still running.
- Clips from any video. Transcribes the audio, asks an AI which moments would hold a stranger's attention, and renders each one with captions in one of three styles.
- Live clipping. Records a stream into a rolling buffer and clips as it happens, using loudness, chat activity and the transcript.
- Workspaces per creator. Channels, clip settings, posting rules and an activity log for each creator you clip.
- Posting. Connect YouTube, Instagram and TikTok in one click, then queue or auto-post finished clips, with spacing and daily caps (see DEPLOY.md for the one-time setup).
- No sign-in on your own computer. It opens straight into the dashboard. On a shared server, switch accounts on
(
PEAKCUT_ACCOUNTS=1) for email or Google sign-in and an admin dashboard. - Your own AI. Gemini, OpenAI, Grok, Groq, OpenRouter, Claude, a local model through Ollama, or anything that speaks the OpenAI chat format. No key at all also works: clips are then picked by loudness.
You need Python 3.10 or newer and ffmpeg. On Windows: winget install Gyan.FFmpeg. On a Mac: brew install ffmpeg.
On Linux: sudo apt install ffmpeg. PeakCut finds it by itself and tells you in the dashboard if it is missing.
Put PeakCut in a short folder that stays put: C:\peakcut on Windows, ~/peakcut on a Mac or Linux. Not inside a
temporary folder or an AI assistant's workspace: those get cleaned up, and on Windows a folder that deep stops the
speech model from loading (Windows can't open files whose full path is over 260 characters). If an AI assistant
installs it for you, it should ask you where to put it first (AGENTS.md tells it to).
On Windows:
git clone <this repository> C:\peakcutcd C:\peakcutOn a Mac or Linux:
git clone <this repository> ~/peakcutcd ~/peakcutpip install -r requirements.txtStart the server:
python -m studio.serverThe first start downloads the speech model by itself (about 480 MB, once).
Open http://127.0.0.1:8765. There is no sign-in: it opens straight into the dashboard. Go to Settings, AI, choose your AI service, paste your key and press Save and test (see Choose your AI).
PeakCut needs a chat model that can return JSON. The easy way is in the dashboard: Settings, AI. Pick the service, paste your key, press Find models to choose from the service's own list, then Save and test. The key is kept in the data folder on your machine and is never shown again.
If you prefer a file (or run a server), the same three things can go in .env. What is saved in the dashboard wins:
PEAKCUT_AI_PROVIDER=gemini
PEAKCUT_AI_MODELS=gemini-2.5-flash
PEAKCUT_AI_KEY=your-key-herePEAKCUT_AI_PROVIDER |
Service | Where to get a key |
|---|---|---|
gemini |
Google Gemini | https://aistudio.google.com/apikey |
openai |
OpenAI | https://platform.openai.com/api-keys |
xai |
Grok (xAI) | https://console.x.ai |
groq |
Groq | https://console.groq.com/keys |
openrouter |
OpenRouter (many models, one key) | https://openrouter.ai/keys |
anthropic |
Claude | https://console.anthropic.com |
ollama |
A model on your own machine, no key | https://ollama.com |
custom |
Any OpenAI-compatible address: also set PEAKCUT_AI_BASE_URL |
PEAKCUT_AI_MODELSis the model id exactly as the service names it. Model names change often, so check the service's current list. Several ids separated by commas are tried in order, which gives you a fallback.- A fast, inexpensive model is the right choice. A one-hour video is read in about four requests.
- You pay the AI service directly for what you use. The AI usage page in the app shows how many requests and tokens each video took.
- Speech to text does not use the AI service. It runs on your own machine, so audio never leaves it.
PeakCut hears English by default. To clip videos in Arabic, add the Arabic speech model once in Settings → Speech models (about 107 MB), or choose Arabic under Options when adding a video and PeakCut offers to download it. Set a creator's language on their Setup page so their live recordings use it too. It understands Modern Standard Arabic and everyday dialects, and the AI turns English words said in between back into English. Captions on the clips are in English.
Everything is set through environment variables, normally in a .env file next to this README.
.env.example lists every setting with a short explanation. The ones that matter first:
| Setting | What it is for |
|---|---|
PEAKCUT_AI_PROVIDER, PEAKCUT_AI_MODELS, PEAKCUT_AI_KEY |
The AI that picks clips and writes titles (or set it in the dashboard) |
PEAKCUT_ACCOUNTS |
1 makes everyone sign in. Use it on any server other people can reach |
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET |
Optional: Sign in with Google, and connecting YouTube |
PEAKCUT_SMTP_*, PEAKCUT_MAIL_FROM |
Optional: confirmation and password emails |
PEAKCUT_DATA |
Where videos, clips and the database are kept (default ./data) |
.env, the secrets/ folder and all user data are ignored by git. Keep keys there, never in the code.
This only matters with accounts switched on (PEAKCUT_ACCOUNTS=1); without them you are the owner and can do
everything. An account is an admin only when the database says so. No email address is special in the code and no web request
can change a role. Do it on the machine that runs PeakCut:
python -m studio.admin grant you@example.comrevoke and list work the same way. The admin dashboard is at /admin and answers "not found" to everyone else.
| Part | What it is |
|---|---|
studio/ |
The server: FastAPI, the clip pipeline, live recording, accounts |
studio/ai.py |
The only file that talks to an AI service. Add a provider here |
web/ |
The pages: landing, sign-in, the app, the admin dashboard (plain HTML, CSS and JavaScript, no build step) |
tests/ |
Tests that run against a throwaway copy of the app |
deploy/, Dockerfile, docker-compose.yml |
One-server deployment with PostgreSQL and automatic HTTPS |
Video work is done with ffmpeg and yt-dlp. Speech is transcribed with sherpa-onnx and NVIDIA's Parakeet
model (English) and FastConformer Arabic model, both released by NVIDIA under the Creative Commons Attribution 4.0
licence (CC-BY-4.0). The Arabic model was converted for sherpa-onnx with tools/export_arabic_model.py.
Data is kept per user: accounts in the database (SQLite on your own computer, PostgreSQL on a server), videos and
clips in each user's own folder. AGENTS.md is a short map of the code for contributors and AI coding
assistants.
Everything PeakCut keeps is in one folder: videos, clips, recordings, creators, settings and the speech model. By
default it is the data folder inside PeakCut. Settings, Storage shows the folder and how much room its drive
has, opens it in one click, and lets you move it to another drive: choose the new folder, and PeakCut moves
everything there the next time it starts. If the drive is nearly full, PeakCut says so and stops taking new videos
until there is room. (PEAKCUT_DATA in .env still works and takes priority.)
PeakCut checks GitHub once a minute while it runs. When a newer version is out, the dashboard shows a notice
within a minute or two (no restart needed), with what changed
and how to update: ask your AI coding assistant to do it (the notice has a sentence to paste), or run
git pull && pip install -r requirements.txt in the PeakCut folder and restart it (with Docker:
git pull && docker compose up -d --build). Your .env, keys and clips are not touched. Only the two commit ids
are sent to GitHub; set PEAKCUT_UPDATE_CHECK=0 to turn the check off.
DEPLOY.md walks through a single server with Docker: the app, PostgreSQL, and Caddy for HTTPS.
docker compose up -d --build- The page asks the server one small question every few seconds. When nothing has changed, the answer comes straight from memory.
- Each person's videos wait in their own line and the lines take turns, so one person cannot hold everyone up.
- The server measures its own processor, memory and disk every second. Extra videos and recordings start only while there is room, wait briefly when it is nearly full, and continue the moment it frees up.
- There are no limits on upload size, video length, recordings, creators or storage unless you set them in
.env. Five videos can be sent at once. Request sizes are always capped, so junk can't fill the memory.
python tests/run.pyThis starts a copy of the app on a spare port with an empty temporary data folder, runs the permission, abuse and queue tests against it, then removes it. It needs no keys and never touches real data.
- With sign-in off, PeakCut listens on this computer only (
127.0.0.1), refuses changes sent by other websites, and only answers to its own address. Do not expose it to the internet that way: turn accounts on. - Passwords are stored only as salted scrypt hashes. Sessions are random tokens stored hashed; cookies are HTTP-only.
- Every route works out who is asking from the session, never from the request, so users only reach their own data.
- Only known video sites are fetched, and text from videos is treated as data when it is shown to the AI.
- Found a problem? Please report it privately through the repository's security advisories instead of opening a public issue.
Clip videos you made, or have permission to use. You are responsible for the rights to what you download, clip and post, and for following the terms of the sites you fetch from and post to.
MIT. Caption fonts in studio/fonts/ (Anton, Archivo Black, Poppins) are under the SIL Open Font
License; the licence texts are in the same folder.