Hermes computer-use: giving each agent its own Linux desktop in a container

What it actually is

Normally a Hermes agent runs commands and reads pages invisibly — you get the result with no view of the process. With this update each agent is given its own desktop on a Linux gateway host, with its own browser and its own saved logins, running in a dedicated Docker container. You drive it from the Hermes terminal as usual and watch the screen respond.

The comparison drawn in the video is to Grok Bot, which uses a similar model of handing each agent a cloud computer. If you have been weighing the two, this narrows a real gap between them.

The command

The CLI surface is a single command with subcommands:

hermes computer-use screen install
hermes computer-use screen start
hermes computer-use screen status
hermes computer-use screen stop

Once it is running you give the agent ordinary instructions in the terminal — "go to example.com and tell me what it is about" — and it navigates on its own screen, reporting back in the session as normal.

The gotcha worth knowing before you debug it

The screen goes black after roughly 30 seconds of inactivity. The container is still running and the agent still responds — it is a blanking behaviour, not a crash. Knowing that in advance saves you restarting a perfectly healthy container.

Why a separate desktop is the safer shape

Two practical benefits, both of which matter more than the novelty. First, the agent is not driving your own desktop — it cannot click through your session, your tabs or your files. Second, because each agent gets its own container image, you can hand it a scheduled task — "do this every day" — and it runs inside its own computer without contending for yours.

That isolation is the point. An agent with a browser and saved logins is a meaningful amount of authority to delegate; keeping it in a container with its own credential store is far better than pointing it at your daily-driver profile. Our Hermes hardening guide covers what else to lock down before running anything unattended.

Key Takeaways

• hermes computer-use screen with install, start, status and stop is the whole control surface.
• Each agent gets its own container, its own browser and its own saved logins — not shared with your desktop.
• A black screen after ~30 seconds idle is expected behaviour, not a failure.
• Scheduled tasks run inside the agent's own computer, which is what makes unattended runs practical.
• Give the setup to a coding agent rather than pasting container commands you have not read.