Why My AI Assistant Lives in Two Places: The Cloud and a Box at Home

<p>Nova is the personal AI assistant I use every day. It doesn't run in one place. Part of it is a cloud tier, and part of it is a home-box tier running local models. It also has a coding agent lane. People ask me why I didn't just pick one and keep it simple. I think the split is the most important design decision in the whole system, so I want to explain why.</p><p>This post is my view on how a personal assistant should be laid out. It isn't a benchmark, and it isn't a guide to what everyone should do.</p><p><strong>The cloud is good at being there all the time.</strong> An assistant you use daily has to be reachable when your laptop is shut, when you're on your phone or when you're away from your desk. Hosted infrastructure does that well. It's always up, you can reach it from anywhere, and it's dull in the way infrastructure should be. If all I wanted was a chat window that answers questions, the cloud would be enough.</p><p>A personal assistant isn't a chat window, though. Over time it collects context about your work, your notes, your projects and your habits. That's where I think the argument for a second tier starts.</p><p><strong>Owning part of the stack is a values decision before it's a technical one.</strong> I care about privacy and about owning my own tools. I don't think everything has to be self-hosted. I think you should be able to choose where things run. A home-box tier with local models gives me that choice. Some work can stay on hardware I control and never pass through anyone else's servers. Even when the cloud version would do the job fine, it matters to me that the local path exists.</p><p><strong>Local models change what's cheap.</strong> I run a bootstrapped business with no outside investors, so I think about cost per call in a way a well-funded team might not. A model on hardware you already own costs about the same whether you call it ten times or ten thousand times. That changes how you design things. You can let background work run more often, retry more freely and try ideas without each one turning into a line on an invoice. I think a lot of agent systems are shaped by API pricing more than their builders realise. A local tier takes some of that pressure off.</p><p><strong>Frontier models are still better at hard reasoning, and I'm not going to pretend otherwise.</strong> Local models have come a long way, but I don't think they replace the best hosted models for every job. Complex reasoning, long-context work and careful code generation are where the stronger models earn their cost. So I don't see two tiers as a compromise. I see each tier doing what it's good at.</p><p>That leads to the real design question, which is where the line goes. My view is that you sort work into three groups:</p><p><strong>Work that has to be reachable.</strong> Anything I'd need from my phone or while I'm out belongs in the cloud tier. Availability matters most here.</p><p><strong>Work that is private, repetitive or happens in the background.</strong> I think this fits the home box best: steady jobs where the quality bar is "good enough and consistent" rather than "the best answer possible", and where I'd rather the data stayed local.</p><p><strong>Work that is hard.</strong> This goes to the strongest model available, wherever it runs. Saving money on the hard problems usually costs more in rework than it saves.</p><p>What I like about this setup is that the routing becomes an explicit decision instead of something hidden inside one vendor's product. When there's only one place to run things, you never have to ask the question. With two, you have to think about each kind of task, and I think that thinking makes the system better.</p><p>There are costs. Two tiers means more moving parts, more to keep in sync and more ways for something to quietly break. A home box doesn't have a data centre's uptime. I don't think those are reasons to avoid the split. They're reasons to design for it: each tier should be able to degrade on its own without taking the whole assistant down.</p><p>Nobody has settled what a personal AI should look like yet. My bet is that the good ones won't live entirely in someone else's cloud or entirely on a box under your desk. They'll run in both places, and the user will decide what goes where. Right now that's how Nova works, and so far it's the setup I trust most.</p>