AnovaX: A Local, Multi-Agent Voice Assistant with LLM Planning, Typed Executors, and Adaptive Recovery

Summary

AnovaX is a desktop voice assistant that runs as a single Python process, keeping voice data local while delegating planning to Gemini. It executes typed agent classes for application control, typing, and browsing within a limited worker pool. The system features a ReAct-style recovery loop and phone-based MJPEG remote access.

Why it matters

The project demonstrates that a readable codebase of just a few thousand lines can coordinate concurrent desktop operations, recover from single-step errors, and be managed remotely without sending raw voice data to the cloud. For privacy-conscious users and researchers prototyping local agent architectures, AnovaX offers a concrete, readable reference application rather than a black-box product. The typed executor and recovery loop pattern provides a practical middle ground between fragile script automation and fully autonomous LLM agents.

Highlights

• Planning is still done in the cloud (Gemini). The 'local-first' claim actually refers to local voice and execution, not local inference.
• MetaAgent's recursive loop is limited to two nesting levels to prevent infinite loops.
• Speculative execution of read-only tools masks Gemini's latency during the recovery phase.
• The Flask server streams the laptop screen to a phone via MJPEG, providing a real-time remote control experience.

Critical take

At 17 KB in size, single-authored, and lacking benchmarks, user studies, or comparisons with existing frameworks (e.g., OS-World, AgentBench, or simple pyautogui scripts), this work reads more like an engineering project report than a research contribution. The 'multi-agent' label also exaggerates what is essentially a typed tool-calling loop with retries. Furthermore, the complete reliance on Gemini for planning undermines the 'local' framework that the title and abstract are built upon.