SharkyClips

Highlights

See it in action

Review a candidate, accept it, render. The clips land in the Clips tab, ready to download.

More screenshots, in both themes, are in docs/screenshots.

How it works

flowchart LR
    subgraph IN[Input]
        A1[Twitch / YouTube / Kick link]
        A2[Video file or OBS folder]
    end
    subgraph LOCAL[Your PC: Python backend]
        B[Ingest<br/>yt-dlp · FFmpeg]
        C[Transcribe<br/>faster-whisper · mlx-whisper]
        D[Chat replay<br/>Twitch · YouTube · Kick]
        E[Signals<br/>loudness · laughter · keywords<br/>speech · game HUD · chat]
        F[Hype Map<br/>rank · snap to words · shape]
        G[Local LLM via Ollama<br/>score · titles · descriptions]
        H[Render 9:16<br/>FFmpeg · captions · bleeps · zoom]
        I[Bleep check<br/>re-listen every clip]
    end
    subgraph OUT[Output]
        J[MP4 / MOV / GIF<br/>SRT · FCPXML · Premiere XML]
        K[YouTube · TikTok · Reels<br/>official APIs]
    end
    A1 --> B
    A2 --> B
    B --> C --> E
    B --> D --> E
    E --> F --> G --> H --> I
    I --> J
    I --> K

The web UI (React) talks to a local FastAPI server on 127.0.0.1. A persistent job queue runs the heavy steps in lanes (GPU, CPU, I/O, network), and a resource governor decides how fast they may run from what you're doing (game in front, fullscreen video, away from the PC). Whisper and the language model never share the GPU. Every step caches its result, so a closed app resumes where it stopped. Details: docs/how-it-works.md.

Tech stack

Layer	Built with
UI	React 18, TypeScript, Vite; 7 locales with a completeness check in CI; dark and light themes; axe-checked accessibility
Server	Python 3.11-3.13, FastAPI + Uvicorn, a persistent multi-lane job queue, a resource governor (psutil, NVML)
Speech	faster-whisper (CUDA / CPU), mlx-whisper (Apple Silicon), a CTC aligner for word timing, sherpa-onnx (laughter, voices)
Analysis	NumPy / SciPy signal scoring, librosa (music mode), OpenCV (facecam), RapidOCR (game HUD)
Language model	Ollama with Qwen 3.5 (4B / 9B, picked from your VRAM), structured JSON output
Video	FFmpeg filter graphs, NVENC / VideoToolbox / QSV / AMF hardware encoding, yt-dlp
Publishing	YouTube Data + Analytics APIs (OAuth + PKCE), TikTok Content Posting API, Instagram Graph API; tokens encrypted with DPAPI / Keychain
Quality	600+ pytest tests on Windows, Ubuntu and macOS; ruff; a Playwright layout check at 9 screen sizes (phone to 1920 px) in 3 languages and both themes; a pixel check that the live preview matches the render
Install	One-click PowerShell (Windows) and shell (macOS) installers, in-app updates via git

This is the beginning of the README. Read the rest at the original. Read the full README ↗