Why I self-host my AI assistant

The question I get most: "why not just use a cloud assistant?" Here's the honest answer — the reasons, the costs, and the trade-offs. 🏠

The short version

I run my AI assistant on my own hardware because three things matter to me more than convenience: privacy, cost, and control. None of them is a niche concern — they're the same reasons people self-host anything.

Privacy: the data stays mine

A cloud assistant is someone else's computer. Anything I send it — my notes, my health data, my work context, my questions — passes through a third party that can log, retain, or use it. For a lot of what I do, that's a non-starter.

Self-hosting means the data never leaves my machine. The model runs locally, the memory stays local, the browser and the pipelines are local. There's no "trust us" — there's nothing to trust, because nothing leaves the box.

The strongest privacy guarantee isn't a policy. It's the absence of a third party.

Cost: predictable, and often lower

Cloud AI is priced per token, and the bill scales with how much you use it. Self-hosting is a fixed cost: the hardware you already own, plus electricity. Once the machine is there, running an assistant on it is nearly free — no per-question meter, no surprise invoice.

For heavy, daily use, that flips the economics. The more I use it, the more self-hosting wins.

Control: I decide how it works

With a self-hosted assistant, I choose the model, the tools, the memory, the guardrails. I can wire it into my own pipelines — a browser, a RAG, a dashboard — without asking anyone's permission or paying for an integration. It's mine to shape.

What it actually costs me

Honest trade-offs, because it's not free:

• Hardware. A small always-on machine. I already had one; it runs the assistant alongside other services.
• Electricity. A few watts of idle draw — negligible on a home bill.
• Maintenance. Self-hosting means occasional upkeep: updates, the odd fix, a watchdog so it stays healthy. That's the price of control.
• Setup effort. It took a day to stand up, not a weekend. The maintenance is the ongoing cost.

When I would use a cloud model

Self-hosting isn't dogma. I use cloud models when the task genuinely needs more than my hardware can run — a very large model, or a one-off heavy job. The rule is simple: sensitive data stays local; the rest can go where it's cheapest and best.

In short

I self-host my assistant for privacy, cost, and control — the data stays mine, the cost is fixed, and I decide how it works. The trade is a little maintenance, which I consider a fair price for owning the thing I rely on daily.

See also: Hindsight + Hermes — giving my assistant a real long-term memory — one piece of the self-hosted stack.