Vision-Language Models (VLMs) are powerful tools capable of interpreting images and videos by generating structured insights such as object detection and contextual understanding, but they often require significant computational resources and pose data privacy risks. To address these challenges, a new method involving Daytona sandboxes and SmolVLM-500M, a compact VLM model, has been developed to safely and efficiently run VLMs on local machines. This approach captures video frames and processes them in an isolated Daytona sandbox using the llama.cpp server, ensuring the host environment remains untouched. By employing NGINX for traffic management, both the frontend and backend can operate on the same port, simplifying the setup. This method allows for fast, interactive processing, generating human-readable summaries from video data without compromising speed or security, and is particularly useful for applications like product demos, content analysis, and safety monitoring.