Mercury – Open-source, local-first agentic harness for Android (github.com)

Mercury has been launched as an innovative, open-source local-first agentic harness specifically designed for Android devices, transforming them into autonomous agent workstations. Building on the Nous Research Hermes Agent project, Mercury enables seamless integration between cloud-based large language model (LLM) reasoning and native Android functionalities. This allows the AI to perform tasks like reading local files, sending emails, and executing commands directly on the device, effectively bridging the gap between cloud capabilities and local computation. The significance of Mercury for the AI/ML community lies in its cutting-edge features, including an embedded Python runtime, device-specific optimization, and a structured memory vault. The agent can execute Python scripts natively, adapt to various Android skins, and automate tasks through a sophisticated system-wide action router that connects LLM directives to Android intents. Enhanced security measures protect API keys with AES-256 encryption, while dynamic tools enable users to create custom skills in real-time. These developments not only enhance user experience but also pave the way for more versatile, responsive AI applications that leverage local resources effectively.