Trying to decide where your always-on AI agent should live, and whether to burn cash on cloud models or run everything local? Let’s make that call quick and painless. I just went deep on a walkthrough from Tina, the ex-Meta data scientist behind Lonely Octopus, who has been running Hermes Agent every single day for over four months. She lays out the exact tradeoffs, the setup steps, and the tricks that separate a smooth build from a frustrating mess. I pulled the whole thing apart so you can pick your path in about five minutes.
The big idea from the creator: a good agent setup is a foundation you keep building on, not a one-time toy. Get the base right and it gets more useful every day. Get it wrong and you fight it forever.
First, pick your criteria
Before choosing anything, the expert says to answer three questions:
- Budget. Are you fine paying for cloud models, or do you want free and private?
- Uptime. Do you need the agent running 24/7?
- Skill comfort. Are you okay in the terminal, or do you want a clean desktop app?
Your answers point straight to the right hardware and the right model. No guessing.
Where your agent lives: the comparison
The original poster runs hers on a Mac Studio (64GB RAM, M4, on 24/7). But she’s clear you do not need that. Here are the four options she breaks down.
- 💻 Dedicated local machine (like a Mac Studio): Pros: powerful local models, fully private, cheapest long-term if you use it heavily. Cons: big upfront cost.
- ☁️ VPS (rented cloud server): Pros: on 24/7, as cheap as 5 or 6 dollars a month, no hardware to buy. Cons: you lean on cloud models, less “yours”.
- Old laptop you already own: Pros: free, where the creator herself started (a 16GB MacBook Pro), just wipe it and leave it on. Cons: limited horsepower for bigger local models.
- Your personal daily computer: Pros: zero extra spend. Cons: risky unless you use Docker to sandbox the agent so it can’t run wild on your files.
Her honest take: heavy users win with a dedicated machine because local models make the running cost basically free. Light or curious users should start on a VPS or an old laptop and upgrade later.
Cloud models vs local models
This is the other fork in the road, and it’s where the money hides.
The general consensus she shares: Sonnet 5 is the best “driver” for Hermes. The catch is you have to use an API key, not a subscription, so it gets expensive fast for daily use. That’s why the creator switched her main driver to an open-source model she runs locally on the Mac Studio. It’s free, private, and she says it handled every demo in the video.
- Choose cloud if: you want top-tier reasoning now and don’t mind the bill.
- Choose local if: you use the agent a lot and want to kill recurring costs. Grab Ollama (the easy route) or llama.cpp (more control and speed, about 30 minutes to set up with the agent’s help).
Either way, if you have no provider yet, she points to the built-in portal subscription that unlocks hundreds of models in one shot.
My recommendation from the walkthrough
If you’re just testing the waters, start cheap: an old laptop or a VPS, plus the model portal. If you already know you’ll live in this thing daily, go local hardware plus an open-source driver. That’s the setup this savvy professional landed on after four months, and it’s the one that scales without a surprise invoice.
The features that make it worth it
The creator walks through five building blocks. These are what you’re actually setting up for:
- Tools: native powers like web search, image generation, and browser automation, managed for you.
- Integrations and MCPs: connect Discord, Telegram, Notebook LM, and more so the agent messages you and pulls research.
- Skills: the headline feature. The agent learns repeatable tasks. She built a “business idea evaluator” once, and now she just types a slash command to score any new idea.
- Cron jobs: scheduled tasks, like a daily news brief turned into a morning audio file.
- Memory: the piece she stresses most.
Getting memory right
This is where most agents fall apart over time. The mind behind the video describes a four-tier stack:
- Core files (memory, user, and agent info) kept short on purpose so nothing bloats.
- Session search that digs through past conversation logs when needed.
- An external memory plugin (she likes Honcho) that quietly learns your patterns.
- An Obsidian “second brain” vault the agent reads from and writes to, so your research and plans stick around.
Her tip: keep the top tiers lean, let the deeper tiers do the heavy lifting. That’s what keeps it fast months in.
Where it ends up
Once the foundation is set, the person who posted it shows a multi-agent system in Discord building small apps in parallel: a Pomodoro timer, a roadmap tracker, all saving to Obsidian. I was genuinely impressed watching two builds run at the same time from a couple of chat messages.
Start with your criteria, pick your machine, pick your model, then layer in skills and memory. Want the exact prompts, model picks, and cron examples she uses? Go watch the full video, it’s worth the time.