AgentArk, a self-hosted AI agent operating system, launches in beta, allowing users to build and run AI agents locally on their machine with built-in safety controls and data privacy.

What happened: AgentArk is a new open-source platform (MIT and Apache 2.0 licensed) that runs as a Docker image on your local machine. It lets you build AI agents from natural language prompts and tools, then deploy them as live apps, scheduled automations, conditional watchers, or chat sessions. The system monitors every action through a feature called Sentinel, distills noisy tool output by 60–90% before it reaches the model, and learns from your corrections and repeated workflows to improve over time.

Why it matters: Unlike cloud-based AI services, AgentArk keeps all your data, conversation history, secrets, and audit trails on your own machine—nothing leaves your device unless you explicitly send it. You can point it at a free local model (Ollama) or bring your own API key to any major provider (Anthropic, OpenAI, Gemini, Groq) and pay only the provider's published rate; AgentArk adds no markup or subscription fee. Every action that touches the outside world goes through a permission gate or approval queue, so the agent cannot run tasks you haven't authorized.

What to watch: AgentArk is currently in beta and not recommended for production use—it can make mistakes and overwrite files in its workspace, though Docker containment prevents access to your host filesystem unless you mount it explicitly. The system requires ~3.1GB for the Docker image and uses ~500MB idle or ~1GB under load. Installation is one command on macOS/Linux or Windows, with configuration available at http://localhost:8990.