What is warden?

Warden is a personal AI assistant that runs directly on your own desktop computer, using local AI models for quick tasks and reaching out to larger cloud models for harder ones, all inside one ongoing conversation. It can browse the internet using your real, already logged in browser, control your mouse and keyboard, run shell commands, and read and send your email. The project's own README opens with a strong safety warning that is worth repeating plainly: Warden runs with the same access as your regular user account, with no sandbox or container separating it from the rest of your system. Because it can execute commands, browse the web, and read messages, a mistake by the AI model, or a malicious webpage or email it processes, could potentially do anything you yourself could do from a terminal. The author explicitly warns against running it on a computer you care about unless you understand and accept that risk, while also noting they personally run it day to day on their own laptop and desktop. The design is built around a small, cheap AI model called the orchestrator, which is the only part you actually talk to. Rather than doing the work itself, the orchestrator figures out what you want and hands the job to one of several specialist AI agents: one that handles heavy research and generation using a large cloud model, one for email and calendar, one for scheduling, one for managing projects, and one for reviewing other agents' work. The orchestrator then checks in on each specialist roughly every thirty seconds, watching whether it is on track, going off course, or stuck in a loop, and can stop and reassign a job if something goes wrong. Finished results land in an internal inbox, which the orchestrator reads through and reports back to you in a single conversational voice, so you only see completed results rather than a stream of behind-the-scenes chatter. It is built with TypeScript and Node.js, uses Ollama to run local models, SQLite for its data, and the Model Context Protocol for connecting extra tools, with a web dashboard for watching what its agents are doing in real time. This project is a fork of an existing project called dockbox.dev. It is released under the MIT license.