Ask the box. Get an answer from the box.

A useful systems assistant has to cross the gap between a sentence and a real operation. General models invented tool names, missed arguments and answered operational questions with generic prose. Even a correct tool call could leave a small model struggling to explain a large raw result. We needed the whole path to work, from intent through execution to an understandable answer.

Hardware made the problem less forgiving. A configuration that felt comfortable on a powerful machine could spend minutes initializing on a small one, or download models that would never fit. We treated that as product behavior to fix, rather than asking an owner to diagnose the assistant before it could diagnose their NAS.

We built Nova for people running their own storage and applications, including creators whose working files live on a NAS and operators looking after containers, shares and virtual machines. They need to understand the actual system in front of them. A plausible general explanation is no help when the question is which service is consuming memory on this particular box.

We built a local assistant with a dedicated tool router and a separate conversational model. An intent classifier sends ordinary conversation directly to the talker. Operational requests go through the router against the documented 41-tool catalogue, then real tool results are converted into a readable response. The installed talker can vary with the machine's capacity.

We added per-user chat history, explicit long-term facts and an audit trail for tool calls and results. An owner can inspect remembered facts, delete one or clear them. User identity travels through the request so memory, chat history and audit attribution agree. The upgrade path also preserved existing chat history rather than abandoning it during the move to separate users.

A storage question can map to get_storage_status without arguments. A container action includes the target and the requested operation. Tools provide the evidence; the talker explains it. The current router receives the full tool catalogue on routed requests. Embedding-based shortlisting is experimental, and we do not describe it as the standard path.

The installer considers memory requirements, including the router that runs alongside a talker. Requests fail early when no model is installed. Hardware detection selects an appropriate local inference engine, while measured profiling checks the engine that actually serves each model. These details let the same product adapt to different machines without presenting every owner with a model-engineering project.

The four-gigabyte Pi exposed assumptions that larger machines concealed. Experimental routing initialization took 161.9 seconds against a 120-second timeout and could never finish. A model could also exhaust memory on an image with no swap. Fixes had to reach existing installations through ordinary updates, including startup repair, rather than living only in newly flashed images.

Memory brought a different failure: a small talker substituted a familiar storage path for the exact path it had been told to remember. We tightened recall instructions and verified the stored value across a fresh chat. Building Nova has meant treating data fidelity, resource limits, migration and measurement as part of the assistant, because a fluent answer is only useful when the machine behind it agrees.