mephisto1122/mevin

Mevin is a self-hosted security-camera dashboard with an AI vision model attached. Once you install it on a computer in your home or office, it can pull video from USB webcams, network IP cameras (the kind that speak RTSP or HTTP), recorded video files, and even the camera in your phone. Each frame gets fed to a local vision model running through Ollama, and the model writes a short sentence describing what it sees. Those sentences stream into a web dashboard in your browser. Nothing is sent to a cloud service. The most distinctive feature is the keyword alert system. The dashboard ships with about forty trigger words like weapon, fire, smoke, fight, intruder, fallen, bleeding, and so on. Whenever the AI description includes one of these words, the matching feed item is flagged with a red ALERT badge, an audio beep plays in the browser, a snapshot is auto-saved, and an optional Telegram message is sent to a bot you set up. The alert keyword list is editable from the settings tab. Getting Mevin running is a short process. You install Ollama from ollama.com, run ollama pull gemma3:4b to download the recommended vision model, then clone the repository, run pip install -r requirements.txt, and start it with python mevin.py. The dashboard opens at http://localhost:5555 and an auto-generated API documentation page sits at /docs. Several other vision models are listed in the README with rough speed and VRAM figures, from a small two-gigabyte moondream model up to a fifteen-gigabyte gemma4:26b for higher quality. The dashboard itself is a single HTML file with vanilla JavaScript, so no build step is needed. It shows a horizontal timeline of events similar to a video-editing tool, with three, six, twelve, and twenty-four hour ranges. A built-in network scanner can find USB cameras and RTSP devices on your LAN, a Guides tab lists the right RTSP URLs for Hikvision, Dahua, Reolink, Tapo, Xiaomi, and ONVIF cameras, and a header strip shows current GPU temperature, VRAM use, and utilisation. Under the hood the backend is FastAPI with uvicorn, and SQLite holds the observations and configuration. Optional bearer-token authentication is available through the MEVIN_TOKEN environment variable. The project is MIT licensed and the README is open to pull requests.