About a year ago I pasted something sensitive into ChatGPT without thinking much about it. Nothing catastrophic, but it made me pause. Financial details, family context, the kind of stuff that I’d never put in a shared Google Doc. The convenience of cloud AI had made me sloppy about what I was sharing and with whom.

That was the privacy wake-up, but privacy alone wasn’t what made me switch. The cost started bothering me more gradually. I was on two separate AI subscriptions, using them inconsistently, and paying whether or not I hit them hard in a given month. When I added it up against what I was actually getting out of each tool, the math felt off. Especially since I had hardware at home that could do a lot of the same work.

The control issue is more subtle but it matters more to me now than the other two. Cloud AI tools have no memory of your environment. Every session starts cold. I’d paste the same project context into different chats, re-explain what I was working on, manually bridge the gap between AI output and action. The AI was helpful but it was disconnected from everything I actually cared about. It couldn’t touch my systems, didn’t know my sites, had no idea what I’d already tried last week.

When I moved to a self-hosted setup, those gaps closed. My local agents have persistent memory. They have tool access. They know the state of my infrastructure. That’s not a privacy choice or a cost choice, it’s a capability choice. The AI became useful in a qualitatively different way when it could actually act on what it knew.

I want to be straight about the tradeoffs though. Running AI at home requires real maintenance. You’re the one responsible when something breaks. Local models are good but they’re not Claude-level on complex reasoning tasks. For anything where I need serious writing quality or complex logic, I’m still sending API calls to Anthropic, just through my own gateway rather than a browser tab. The cost of those calls is much lower than a flat subscription when usage varies month to month.

The data question is real too. When you use cloud AI in a browser, you’re trusting that company’s data policies and their security posture. When you run it locally, you’re trusting yourself. I’d argue most homelab people are pretty motivated to keep their own systems clean, but it’s not zero risk; it’s different risk. One piece I’ve added to my own setup is a YubiKey 5 NFC on accounts that touch the server and the WordPress admin logins. When you’re the one responsible for your own infrastructure, hardware 2FA is an easy layer to add.

What I’ve settled on is a hybrid. Routine tasks, anything involving personal data about my family, infrastructure queries, content management: all local, all through my own agents. Tasks that genuinely need the strongest available model: API calls to cloud providers, but with me controlling what data gets sent and when. I’m not sending my full server state to an LLM; I’m sending a narrow, deliberate query.

The other thing I stopped doing: using AI as a glorified search engine. The cloud tools train you to ask one-off questions. Once your AI has context and tools, you start thinking in workflows instead. That change in how I frame tasks is probably more valuable than any of the technical decisions.

If you’re on the fence about this, I’d say start by auditing what you’re actually doing in cloud AI sessions. How much of it is personal data you’d be uncomfortable with on a shared doc? How often are you re-explaining the same context? That audit will tell you whether the switch is worth it for your situation.

Hardware linked in this post:

Affiliate disclosure: Some links in this post are Amazon affiliate links. If you buy through them, I get a small commission at no cost to you. It helps keep the lights on here.