PeakCut is available as a public, MIT-licensed project for running long-video and live-stream clipping on a developer’s own computer or server. It accepts uploads or links from YouTube, Twitch and Kick, then renders selected moments as 9:16 clips with word-by-word captions and hooks. The practical consequence is that developers can operate the server, media storage and speech-recognition stages themselves rather than depend on a managed clipping application. The repository documents the pipeline and licence.
AI selection is optional
With an AI service configured, PeakCut transcribes a video and asks a chat model that can return JSON to identify moments likely to hold attention. Its documented choices include Gemini, OpenAI, Grok, Groq, OpenRouter, Claude, a local model through Ollama and custom OpenAI-compatible endpoints. Multiple model IDs can be configured as fallbacks. Without a key, PeakCut still operates but selects clips by loudness instead. PeakCut describes both model-backed and keyless selection.
Operators can configure the provider, model and key through the dashboard or environment variables. A key saved through the dashboard is kept in PeakCut’s local data folder and is not displayed again. Users pay a selected hosted AI service directly, while an in-app usage page reports requests and tokens consumed for each video. The repository explains its key handling and usage reporting.
What stays on the host
Speech recognition runs on the operator’s machine using sherpa-onnx. The project uses NVIDIA’s Parakeet model for English and offers a separate FastConformer model for Arabic; according to the README, audio does not leave the machine for this transcription stage. Video processing uses ffmpeg and yt-dlp, while the server is built with FastAPI and the web interface uses plain HTML, CSS and JavaScript. PeakCut documents these implementation components.
The local quick start requires Python 3.10 or newer and ffmpeg. PeakCut starts at 127.0.0.1:8765, initially opens without sign-in and downloads an English speech model of about 480 MB on first use. Its files—including videos, clips, recordings, settings and speech models—are kept under a configurable data directory. Installation and storage behavior are specified in the repository.
For live streams, PeakCut records into a rolling buffer and evaluates loudness, chat activity and the transcript while the stream remains active. Creator workspaces hold channels, clipping settings, posting rules and activity logs. Finished clips can be queued or automatically posted to connected YouTube, Instagram and TikTok accounts, subject to configured spacing and daily limits. The project describes its live-clipping and posting features.
Local and shared deployments differ
With sign-in disabled, PeakCut binds to 127.0.0.1 and warns operators not to expose that mode to the internet. Setting PEAKCUT_ACCOUNTS=1 enables account mode for a shared server; Google sign-in and YouTube connection additionally require Google client credentials. The supplied single-server deployment uses Docker, PostgreSQL and Caddy for HTTPS. PeakCut’s deployment documentation distinguishes these configurations.
Analysis: control shifts responsibility
Analysis: PeakCut’s significant design choice is the separation between local media processing and configurable clip selection. Ranking can use a hosted provider, local Ollama model, custom compatible endpoint or no model key, so “bring your own AI” is an option rather than a hard dependency. Self-hosting also does not make every configured path local: hosted models, source platforms and posting integrations remain external when an operator chooses them. Those boundaries follow from the documented provider, ingestion and posting options.
That flexibility transfers operational work to the person running the service. The operator is responsible for storage capacity, credentials, account configuration and safe internet exposure. PeakCut includes disk-space checks, per-user storage separation, resource-aware queuing and deployment instructions, but the repository provides no universal processing-time figure for a given machine. Capacity planning therefore remains installation-specific. The repository describes its storage, queueing and deployment controls.
Content rights remain another explicit boundary. PeakCut instructs users to clip material they created or have permission to use and makes them responsible for complying with the terms of the services from which they fetch and to which they post. That responsibility appears in the project’s usage guidance.

