Kiffos is a native macOS application that provides a local artificial intelligence assistant designed to operate entirely on the user's machine. It functions as a personal operating system layer, running as a GPU-accelerated app on Apple Silicon (M1 through M4) with macOS 11 or later. The application processes all data locally, requiring no cloud connectivity, no user account, and no subscription. Every action taken by the AI agent is subject to a permission verifier that checks system calls against defined policies before execution, with sensitive operations requiring explicit user approval. The interface is currently in French, with an English version planned for future release. The application includes several core capabilities. It supports voice commands for natural language interaction, document creation and export to PDF or Markdown, email management via IMAP and SMTP protocols, file browsing, and an application launcher. The AI model is selected based on the available RAM, with options for 3 billion, 7 billion, or 14 billion parameter models. Kiffos also includes a local network monitoring feature called EHA (Endpoint Host Anomaly) defense, which maps traffic and flags abnormal patterns. A Pro version adds alerts, automatic responses, and peer-to-peer mesh networking. The application can optionally connect to remote models via API keys compatible with vLLM or OpenAI endpoints. Typical use cases involve interacting with the assistant through voice or text to perform tasks such as opening applications, reading and replying to emails, composing structured documents, and managing files. The agent executes these actions under the governance of local policies, ensuring that no data leaves the machine. The application works offline and does not require an internet connection for its core functions. Users can also use the cockpit interface to browse and launch applications without leaving the Kiffos environment. Technically, Kiffos runs as a native app with Metal-accelerated inference, requiring at least 8 gigabytes of RAM. It is designed for Apple Silicon processors and macOS 11 or newer. The download, license, and update mechanisms function globally, independent of language support. The long-term development roadmap includes a separate research project to implement the same sovereign core on a bare-metal Rust microkernel at Ring-0, but this is distinct from the current native application.