Launching today
Hot Desk
Local Linux desktops for AI agents, with human takeover
4 followers
Local Linux desktops for AI agents, with human takeover
4 followers
Hot Desk gives each AI agent its own local Linux desktop. The home directory and Chromium profile persist, so a site you sign into once stays signed in for later sessions. Your agent connects over MCP and receives the browser, computer and workspace tools directly, with no tokens and no model API key in the config. When it gets stuck on a login or a CAPTCHA, you open the viewer, take over the same desktop, then hand control back. Runs on Docker on your own machine. MIT licensed.
Savvy
I kept handing browser tools to coding agents and watching them stall in the same place: a login screen. The session ended, the cookies went with it, and the next run hit the sign-in wall again or quietly made the answer up.
Hot Desk is what I built to stop that. Every desktop is a Docker container with its own home directory and Chromium profile, both persistent, so signing in once holds. You set one up once:
hotdesk up
hotdesk open research
Sign in to X and Google in that Chromium window, then hit "Return control to agents". After that the whole prompt is:
"Use the Research Hot Desk desktop for this task. Research browser agents on X and save a Google Doc with 3 sources and 3 product takeaways."
The agent arrives already signed in and does the rest. It connects over MCP to that one workspace and receives the browser, computer and workspace tool schemas at connection time, along with the handoff and clone rules. No tokens and no model API key go in that config, and the manager binds to 127.0.0.1.
When it gets stuck, you open the viewer, take over the same desktop, do the human part, and hand control back. The demo above is Claude Code running exactly that prompt, with me stepping in mid-task.
Where it actually is: early, no stable release, macOS or Linux with a local Docker daemon. Desktops are containers sharing the host kernel, not VMs, so give an agent the same trust you would give it on your own machine.
Two things I would like your read on:
1. How much should an agent be allowed to do inside a profile that is already signed in as you?
2. One long-lived desktop per task, or a fresh clone from a checkpoint per run?
Code and issues: https://github.com/jamalavedra/h...