Kubi
Always-on Linux desktops for AI agents.
Kubi gives your AI agent its own Linux desktop that stays on. We call it a screen. Your agent looks at it, clicks and types like a person, runs commands and moves files. Your files, logins and open apps are still there tomorrow. A person can watch live and take over at any time.
Three ways in
| You use | Start here | Good for |
|---|---|---|
| An agent app (Claude Code, Codex, Cursor, claude.ai) | Connect your agent | Giving the agent you already use a desktop, in two minutes |
| Your own code | Quickstart with the TypeScript SDK | Agents you build and run yourself |
| Any language | API reference | Plain HTTP and JSON, curl works |
All three run on the same API, so anything one can do, the others can too.
What an agent can do on a screen
- Look. Take a screenshot and get the image back with an id. Screenshots
- Act. Click, type, press keys, scroll, drag. Each action can return the screen after it, so one call is one step. Actions
- Run commands. Short ones answer inline, long ones return an id to check later. Commands and files
- Move files. Upload, download, list folders.
- Share it. You can watch live or take control at any time, and your agent's calls are refused until you're done. People and agents
- Look back. Every call is kept with the screenshot it returned for 30 days. Billing and history
How a screen works
Each screen is a full Linux desktop (Debian 13 in Ubuntu's look) in its own virtual machine, with browsers and the usual tools. It runs until you stop it. A persistent screen is billed monthly, running or stopped; a disposable one is paid from credit for the time it runs and can end on its own. Save any screen as a template to start new ones from. Screens
Agents that read docs: everything here is also at /llms.txt, and every page has a Markdown copy at the same address plus .md.