AnovaX: Local Multi-Agent Voice Assistant with LLM Planning

Traditional desktop voice assistants often rely on cloud processing, sending user audio off-device and offering a limited, fixed set of functionalities. A new project, AnovaX, challenges this paradigm by introducing a local-first, multi-agent voice assistant designed to run entirely on the user's machine, treating the desktop itself as its primary interface for actions. AnovaX is built around a single Python process that integrates several key components: a wake-word detector, a speech pipeline, and an LLM planner (Gemini) that generates structured JSON plans for tool execution. It also includes a safety layer, a multi-agent orchestrator that dispatches tasks to specialized child agents (e.g., AppAgent, TypingAgent, BrowserAgent) on a thread pool, and an adaptive recovery loop to handle failures. The system supports recursive sub-goal delegation via a MetaAgent and uses speculative execution to mask LLM latency. A companion Flask server enables remote control from a phone, mirroring agent events and streaming the laptop screen. AnovaX demonstrates that a relatively compact, local assistant can perform complex tasks like opening apps, typing, searching, coordinating actions, and recovering from errors, all without cloud dependency or direct LLM keyboard access.