Nous Research adds live screens and human takeover to Hermes agents

Why it matters

Live viewing and takeover make remote agents easier to supervise through login and CAPTCHA steps, but Nous' docs say separate bot screens do not isolate bots from the shared host, files or network.

Nous Research added a Bot Screen feature to Hermes Desktop on September 23rd, letting users watch an agent operate a browser or terminal on a remote machine and take over its screen when a task needs human input. The company described the feature in a post on X, alongside a walkthrough by tonbi (@tonbistudio).

https://x.com/NousResearch/status/2102873234819698877

The feature addresses a practical gap in computer-using agents: a bot can reach a login prompt or CAPTCHA and stop, leaving a person to finish the step. In its documentation, Nous says a user can take control of the screen, complete a login or other action, then hand control back so the agent can continue in the same session. The screen runs on the gateway host, so an agent can keep working there after the user closes the desktop app or turns off their laptop.

The implementation is aimed at remote Linux machines, including cloud virtual machines. Each Hermes profile gets its own screen, browser profile and cookies, according to the docs. The user can open the screen from the bot view or sessions sidebar, watch a live preview, and bring the screen forward when a bot starts using its computer or browser tools. The feature is off by default; users have to start a screen or enable automatic startup.

That makes the takeover feature a handoff mechanism, not a security boundary. Nous explicitly warns that bots on the same host share the operating-system user, files and network. A process running as that user, including another bot or Hermes' terminal tool, can access the screen and browser profile outside the desktop pane. The docs advise users to avoid typing secrets into a bot they do not trust. When a human takes control, Hermes blocks the bot's computer-use and browser tools, but the documentation describes that as a tool-level restriction rather than operating-system isolation.

Running the screen also has deployment costs. The gateway needs Linux, TigerVNC and Xfce components, plus computer-use support for the bot. Nous estimates that an open screen with a browser uses about 1.1 to 1.5 GB of memory; the screen without a browser is substantially lighter. The screen can stop after an idle period and restart when needed, but teams running agents on small machines will need to account for the additional memory and host setup.

The update puts an operator directly in the agent's work surface instead of asking them to rely only on a chat transcript or a report after a run. That is useful for tasks involving authentication or other steps that require a person, while keeping the agent's session on the remote host. It also exposes the operational tradeoff: the same shared host that makes a continuing session convenient means a screen per bot should not be mistaken for isolation between bots.

Nous Research was founded in 2023 by Jeffrey Quesnelle, Karan Malhotra, Ryan Teknium and Shivani Mitra, according to TechCrunch's reporting on the company. The Bot Screen announcement gives no performance or adoption figures. Its release is a product-level addition to Hermes Desktop, which Nous describes in the project repository as a desktop interface for the same agent available through its command-line and gateway tools.