Hermes Agent Launches macOS Computer Control with 95% Reduction in Token Usage

According to monitoring by Beating, Nous Research’s Hermes Agent has officially launched its Computer Use functionality on macOS. This feature directly competes with OpenAI Codex’s “background control” capability. Under the hood, it integrates with cua-driver—an open-source driver previously reported on in this channel—that reverse-engineers Apple’s proprietary APIs to issue direct operational commands to target processes. This means that when the Agent is searching emails or typing code in the background, the user’s physical mouse won’t jump erratically, and the current window focus won’t be stolen, enabling seamless, non-disruptive collaboration between human and agent on the same machine. Since computer control heavily relies on continuous screenshots, token costs can skyrocket. To address this, Hermes implements a four-layer context compression mechanism at the framework level: forcibly eliminating redundant visuals, retaining only the most recent three screenshots in memory, and coordinating server-side cleanup of old caches. Official estimates show that, when executing 20 consecutive actions at a resolution of 1568×900, context token consumption drops dramatically from approximately 600,000 tokens to around 30,000.