Why Bother? Give Your Local AI a Job: Meet the AI Butler
A local chatbot is just a more private version of what you already have. Wrap a model in an agent and it becomes a butler: one that remembers you, acts on your behalf, and answers to nobody else.
So far I’ve sold you on a model running on your own computer. Free, private, yours. But if all you do is open a window and chat with it, you might fairly ask: so what? The frontier chatbots do that already, and they’re smarter.
Fair. A local model on its own is an engine sitting on a workbench. Impressive, but it doesn’t take you anywhere until you put it in a car and give it somewhere to go. This post is about the car: turning your local model from a chatbot you visit into an assistant that actually works for you.
From chatbot to butler
The chatbots we all use are reactive. You ask, they answer, and then they forget you existed. Useful, but passive. You’re always the one starting the conversation and doing something with the reply.
What changes the game is wrapping your local model in something called an agent. An agent is a layer of software that gives the model three things a bare chatbot lacks: memory that lasts, the ability to actually do things rather than just talk, and the patience to work through a multi-step job on its own.
The analogy I keep coming back to is the difference between a clever stranger and a butler. A clever stranger can answer any question you put to them brilliantly, but they don’t know you, they can’t act on your behalf, and tomorrow they won’t remember you met. A butler is different. A butler remembers how you take your coffee, knows where things are kept, and can be handed a task (“sort out the morning post”) and just get on with it.
That’s the leap. The model is the cleverness. The agent turns the cleverness into a butler.
What a butler actually does for you
Made concrete, here’s the kind of thing this unlocks once a local model has an agent wrapped around it:
It remembers across days, not just within one chat. Tell it something on Monday and it still knows on Friday. The frustrating goldfish memory of normal chatbots goes away.
It can reach into your tools (your files, your calendar, your messages) and act, rather than just describing what you should do. Ask it to pull together your morning brief and it can actually go and assemble one.
It can be handed a job and left to work. Not every reply needs you sitting there. You set it going and come back to the result, which is exactly where the patient-but-slower local model earns its keep.
And because the whole thing runs on your own machine, all of that memory and all of that access stays with you. A butler who works in your house and gossips to nobody.
The agent I chose, and the one I didn’t
When I went looking for the software to do this, two open options dominated, and they reflect genuinely different philosophies. It’s worth understanding the choice, because picking the wrong one means fighting the tool instead of using it.
The popular one is called OpenClaw. It’s the crowd favourite by raw numbers, and its strength is reach: it connects your assistant to just about every messaging app you already use (Telegram, Discord, Slack, WhatsApp, Signal, iMessage and twenty-odd others) so you can talk to it from anywhere. Think of it as what Siri should have been: open, extensible, reachable from everywhere. If your dream is one assistant you can poke from your phone on any platform, that’s its sweet spot.
The one I chose is called Hermes, built by a respected AI research group called Nous Research. It made a different bet, and it’s the bet that won me over. Where OpenClaw chased breadth of integration, Hermes focused on depth of learning. Its whole architecture is built around a loop that makes the agent better the more you use it. OpenClaw’s skills are static files you write and maintain; Hermes creates its own skills as it works, refines them during use, and compounds them across sessions.
In butler terms: OpenClaw is a butler you can phone from anywhere. Hermes is a butler who actually learns your household and gets better at the job every week. Its tagline is “the agent that grows with you,” and that’s exactly the quality I wanted.
Two more things tipped it for me. First, it’s a natural fit for local models specifically. It works with any model offering a long enough memory, including local ones, and it’s free to self-host with your only real cost being optional hosting of a few dollars a month. Second, the safety posture is more careful by default: Hermes ships with built-in scanning for malicious instructions and credential filtering, whereas OpenClaw’s rapid growth has come with a string of security issues that needed patching. For something that’s going to have access to my files and memory, careful-by-default mattered.
One honest caveat. Hermes is the newer, smaller project and it leans toward the research-and-tinkerer crowd. It also wants a Unix-like environment, so on Windows you run it through a compatibility layer called WSL2. None of that is a barrier for a weekend project, but it’s the kind of thing worth knowing going in. If you’d rather have the broadest reach and the biggest community, OpenClaw is the reasonable other choice. And helpfully, you can migrate from one to the other with a single command, so the decision isn’t a one-way door.
Where this is heading
I won’t pretend that wiring up an agent is as simple as the chat-in-a-window setup from the next couple of posts. It’s the more advanced end of the weekend project, and I’d genuinely suggest getting comfortable just chatting with a local model first before you give it a job.
But I wanted to plant the flag early, because it’s the why. The reason to put a model on your own machine isn’t to have a slightly more private chatbot. It’s to eventually have a capable, private assistant that remembers you, works for you, and answers to nobody else. That’s a thing worth owning.
Next, we get practical: figuring out which model your computer can actually run, and the surprising trade-off I found between fast and smart.