What is a good desktop AI agent that is private and open source?
I want an agent that can actually work on my computer, clicking through apps, filling forms and moving files, but I would rather it be open source and not send everything on my screen to a company I can't inspect. What options are people using right now?
Most desktop agents are two pieces: a model that decides what to do, and a computer-use layer that takes screenshots and sends clicks and keystrokes. Open-source options usually cover the second piece, often as an MCP server, so you can plug it into the agent client you already use.
Be clear about what open source buys you. You can read exactly what the tool does with your screen, but the screenshots still go to whichever model you connect. For real privacy you need a local model as well, and local models are noticeably weaker at reading screens and planning long sequences of clicks.
Desktop control is the riskiest access you can give an agent: it can see whatever is on screen and act as you in any logged-in app. Use a separate user account or a virtual machine, close password managers and banking tabs, and keep confirmation on for anything that sends, buys or deletes.
For jobs inside a browser, a browser-only agent is usually more reliable and easier to fence in than full desktop control.
Listings mentioned
- Auten MCP · agent by auten-aiMIT-licensed computer-use MCP server that gives any MCP client control of a real screen on Mac, Windows or Linux.
- Interceptor · skill by danielmiesslerDrives a real Chrome or Brave session plus macOS computer use, useful when the work lives mostly in the browser.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.