Hermes Agent Launches macOS Computer Control Feature, Screenshot Token Consumption Reduced by 95%

According to Conduct Beat monitoring, Nous Research's Hermes Agent has officially launched the macOS Computer Use feature.

This feature directly mirrors the OpenAI Codex's "background control" in terms of user experience. It integrates the open-source driver, cua-driver, as previously reported in this channel, to interact with the target process through reverse-engineering Apple's private API to directly issue operation commands. This means that when the Agent is checking emails or coding in the background, the user's physical mouse cursor will not move erratically, and the current window focus will not be stolen, enabling human-machine collaboration on the same computer without interference.

Due to the heavy reliance of computer control on continuous screenshots, the Token bill often inflates rapidly. To address this, Hermes has implemented a set of four-tier context compression mechanisms at the framework level: forcefully removing redundant screens, only allowing the model to retain the last 3 screenshots, and coordinating with the server to clear old caches. According to official calculations, performing 20 consecutive steps at a resolution of 1568×900, context consumption can drop from approximately 600,000 Tokens to around 30,000.