Nous Research adds CUA computer use to Hermes Agent for desktop control
Nous Research added early computer-use support to Hermes Agent through CUA, enabling background desktop control without taking over keyboard, mouse, or screen input. The feature opens computer use to local or alternative models instead of tying the workflow to frontier-only modes.

TL;DR
- Teknium's preview post says Hermes Agent now has an early computer-use mode powered by CUA, and that it works with "any model" rather than a frontier-only special mode.
- The main usability claim in Teknium's thread is that Hermes runs computer use in the background, so it does not seize your keyboard, mouse, or screen while the agent operates.
- NousResearch's demo post shipped the feature with a short video and linked docs, which now document a dedicated computer-use page.
- Hermes already had agent infrastructure around this launch, including the v0.13.0 collaboration board and
/goalloop that WesRoth's release summary highlighted, plus provider rotation through credential pools that Teknium's credential-pool tip surfaced. - The wider Hermes stack was already leaning toward model portability, with mervenoyann's Hugging Face post adding local-app support and traces, and OpenRouter's docs screenshot showing auxiliary-task routing through Pareto Code.
You can watch NousResearch's demo clip, jump straight to the computer-use docs, and trace how Hermes was already being wired for multi-provider operation through OpenRouter routing docs and Hugging Face local-app support. The interesting part is not just desktop control. It is that Nous is trying to make computer use ride on the same portable agent stack as local models, auxiliary routers, and trace tooling.
Computer use
The launch claim is simple: Hermes Agent can now control a computer through CUA. NousResearch's post framed it as "computer use with any model," while Teknium's preview post repeated that the feature is built into the latest Hermes Agent and opens desktop control to models outside the usual frontier-only computer-use products.
The official docs make this a first-class feature, not a one-off demo, via a dedicated computer-use documentation page. trycua's repost and Teknium's follow-up reply mostly functioned as confirmation that CUA is the integration layer behind the release.
Background control
The most concrete product detail came from Teknium's preview post, which said Hermes does not take over the PC and instead works "entirely in the background" while the user keeps control of keyboard, mouse, and screen.
That is a different interaction model from the remote-browser or locked-session feel common in early computer-use demos. Nous is pitching concurrent use of the same machine, not a handoff.
Agent plumbing already in place
Computer use landed on top of a broader Hermes push toward persistence and reliability. WesRoth's release summary said v0.13.0 added a durable collaboration board, a /goal command to keep the agent pinned to a target across turns, Checkpoints v2, auto-resume, and eight closed P0 security issues.
Two pieces matter for this rollout:
- Multi-agent task handling, via the collaboration board in WesRoth's release summary, gives Hermes somewhere to coordinate longer-running work.
- Provider failover, via credential pools that Teknium's credential-pool tip highlighted, lets one provider rotate across multiple API keys for stability.
That does not prove how every computer-use task is executed internally. It does show the feature arrived inside a harness already optimized for retries, resumption, and provider churn.
Local and routed model support
The "any model" pitch also rests on the rest of the Hermes ecosystem. mervenoyann's Hugging Face post said Hugging Face added Hermes Agent to local apps for compatible GGUF and MLX models, and added native Hermes trace visualization on the Hub.
Meanwhile, OpenRouter's docs screenshot pointed to a config path where Hermes can send auxiliary tasks through OpenRouter's Pareto Code router. The linked docs show provider ordering, throughput sorting, provider allow or ignore lists, and a min_coding_score knob:
provider: openroutermodel: openrouter/pareto-code- ordered provider preferences such as Anthropic then Google
- routing by
throughput, with docs notingpriceorlatencyas alternatives - provider include or exclude controls
- a Pareto router
min_coding_scorethreshold
That is the deeper reveal in this launch. Computer use is new, but Hermes was already being assembled as a portable agent layer that can swap models, route subtasks, and run locally.