It was 11pm and I needed to check a server from my phone, no laptop in reach. I opened an improvised SSH terminal, connected to the VPS, and there it was: the same Claude conversation I'd left half-finished at my desk that afternoon, as if I'd never left. That small, almost forgettable moment ended up changing how we think about AI infrastructure at Guatemalia.
When someone says "let's run AI on our servers," the instinctive reaction is to reach for a calculator: how much RAM do I need, what GPU do I buy, how much will it cost to scale this? It's the same reflex as someone who thinks they need a bigger kitchen to cook faster. Sometimes you do. But sometimes what you actually need is a better menu, not more stoves.
We run Claude Code directly on our VPS. Connecting over SSH — from PuTTY, from another terminal, from wherever — we found that the working session doesn't live in the window you opened. It lives in the process running on the server, like a paused conversation you pick back up exactly where you left it, no matter which door you walk in through.
But persistence was only the surface. What was actually worth examining was where the real work happens.
Here's the part that changed our workflow the most: the SSH connection is only needed the first time, to open the session. After that, Claude keeps running server-side, with the full project context already loaded — it isn't a terminal that switches off when you close the window.
That means we don't have to repeat the usual ritual — logging in, finding the folder, remembering which command goes where — every time we need something checked or executed on the server. Claude is already there, ready, with memory of what was done before. You just tell it what's needed, from anywhere, and it takes care of running it remotely, directly against the live system.
Think of it this way: our VPS works like the receptionist of an office building, not the engineering team on the 20th floor. Its job is light — note who came in, what they asked for, where the conversation left off. It stores context: text, file references, command output. None of that needs a GPU or industrial-grade RAM.
The heavy lifting — the part that actually "thinks" — happens on Anthropic's infrastructure, through the subscription we're already paying for. We're not "sharing RAM between servers" like some kind of sci-fi distributed computing setup. We're doing something simpler and more elegant: separating who takes the call from who solves the problem.
The benefit shows up on two fronts.
For small teams like ours, this replaces the wrong question — "how much hardware do I need to do AI?" — with a better one: "what part of this work actually needs to live on my server, and what part can I hand off to a cloud I'm already paying for?"
In our case, the answer was almost embarrassingly simple: almost all of the weight can be handed off. And what's left is local capacity free for the one thing that actually matters — serving the people who use Guatemalia better.
Sometimes the best optimization isn't running faster. It's letting go of things you never should have been carrying in the first place.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG and orchestrators for companies across Guatemala and Latin America. Reach out for a consultation.
Contact Carlos Montiel