AnovaX voice assistant runs locally on user's computer

AnovaX, a new voice assistant described in a preprint on arXiv, runs entirely locally on a user's computer using a single Python process and Gemini-generated JSON plans, avoiding cloud-orchestration frameworks. The architecture uses typed executor agents with bounded thread pools, speculative execution of read-only tools to hide LLM latency, and a localized ReAct recovery loop capped at two recursive planning levels, demonstrating reliable desktop automation with local safety whitelists.

arXiv https://arxiv.org/abs/2607.15367 AnovaX voice assistant runs locally on user's computer Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated. AnovaX demonstrates that robust desktop automation can run entirely in a single local Python process, using a bounded thread pool of typed executor agents directed by Gemini-generated JSON plans. To maintain responsiveness, the architecture hides LLM planning latency by speculatively executing read-only tools while a localized ReAct recovery loop handles single-step failures within a hard limit of two recursive planning levels. For production engineers, this proves you can bypass complex cloud-orchestration frameworks and ship reliable, self-recovering OS agents using lightweight thread locks, local safety whitelists, and structured JSON planning. The whole architecture is a few-thousand-line single Python process: LLM Gemini does planning only—emitting JSON tool-call plans—while typed per-tool agent classes each with own timeout, retry, resource locks do the actual desktop actuation, gated by a whitelist/denylist and a ReAct recovery loop that speculatively runs read-only tools to hide LLM latency. The pattern worth stealing: keep the model out of the actuation path entirely, wrap each executor as a typed agent with bounded concurrency and its own retry/lock policy, and cap recursive self-delegation here at two levels so plan expansion stays predictable and legible.