Setting up your Hermes agent on a €3.60/month VPS: the complete zero-to-live tutorial (even if you’re not a dev)

You want your own AI agent, the kind that runs 24/7 on your business without asking you a thousand questions — but the prospect of “renting a server” and “generating an SSH key” makes you want to put it off until next Monday? That’s exactly why I filmed this tutorial. The promise of this article: you’ll walk away with a working Hermes agent on your own VPS, installed by your AI, not by you. Your actual job is to answer its questions.

One honest reminder before we start: Hermes is only worth it if you already know why you’re creating it. If you install it “just to play,” you’ll find it useless — and you’ll be right. If you’ve already identified repetitive tasks you’d happily delegate without paying someone €3,000/month for it, then this setup makes sense. The full video runs 37 minutes; this article covers everything, in order, with the real numbers.

The full video (37 min, the setup filmed for real)

If you’d rather watch the process live — including my hesitations and the moments where I get lost in my own screens — here’s the original video:

▶

Step 1: the VPS — your playground, not a bloated machine

What’s a VPS? A small slice of a computer in a data center, rented by the hour, week or month. It runs 24/7, it’s connected to the internet, and you don’t physically own it — exactly what you need for an agent that has to live without you.

In the video, I use Contabo, but Hetzner or another provider will do just fine. The only advice that matters at the start: begin with the cheapest option. You can always upgrade; the reverse is more painful.

Two profiles, two choices:

• Basic VPS: 100 GB SSD, 8 GB RAM, 4 cores. More than enough to get started, and even to do a lot more than “get started”.
• My setup (Storage VPS 10): €3.60/month, real price. Slower, a bit less RAM, but 300 GB of SSD + 150 GB of NVMe. I’m an SEO cheapskate and above all I never want to run out of space. An agent working in the background doesn’t need Formula 1 speed.

Three settings you shouldn’t miss when ordering: Ubuntu image, server in Europe, and you pay. You receive two things by email: the IP address (the address of your online home) and the root password. You’ll need them right after — and never again afterwards.

Step 2: the SSH key — your only real technical step

Open the AI you already use — Claude, Antigravity, Codex, whatever — and tell it: “I want to create an SSH key.” It will handle everything, and this is already where the tutorial’s philosophy kicks in: your AI uses the server, not you.

An SSH key is a file split into two parts: a public key that your computer will drop onto the VPS, and a private key that never leaves your machine. Concretely: your AI generates the key, gives you a command to copy-paste, you run it, the VPS asks for the root password received by email to verify it’s really you, and the file is deposited. That’s it.

From that moment on, you can close your VPS provider’s interface and forget about it. You won’t go back — your AI accesses it directly over SSH.

Step 3: security — because your machine is online 24/7

An online server is constantly exposed, and nobody watches it 24/7. The first instruction to give your AI: secure the server, and install Fail2ban. It’s protection against brute-force attacks — usually the AI installs it on its own during preparation, but ask for it explicitly: it doesn’t hurt, and it forces you to understand what runs on your machine.

Then, still in the same AI conversation: “I have a VPS, I have SSH access, I want to install Hermes Agent. What are the main steps?” And you let it work. You don’t need to know what Docker or systemd are at this stage — the first agent installed is the system agent, let it install itself. If you don’t know how to do something, ask: we live in a world where asking your AI costs nothing, and explaining out loud personally gave me a better understanding than when I installed everything myself.

Step 4: the LLM — giving your agent a brain

Hermes without an LLM is a body without a brain. It needs a way to think, and you choose which one:

• Starting from zero? Get a ChatGPT Pro subscription ($20/month) and connect it: it’s the simplest, and you also get image generation as a bonus, which is very handy.
• Already have Claude? Connect Claude to your existing subscription, same logic.
• Want something cheap on usage? A DeepSeek API key: you provide the API address + the key, and you’re done.
• My choice: GLM 5.3 Flash, a subscription at about $15/month. Why? Because I don’t ask Hermes to be a genius. I ask it to be consistent, reliable, to work in the background without asking me ten thousand questions. GLM 5.3 Flash is more than enough for that.

Worth noting: whichever option you pick, the installation is driven from your desktop AI (Claude, Antigravity, Codex). You can actually chat with Hermes through it — but that’s not the point: the everyday work interface comes in the next step.

Bonus option, the fallback: if service continuity matters to you, configure a second default LLM. Example: ChatGPT as primary, and a GLM Flash or DeepSeek API key as backup. If the first one goes down, Hermes switches over and keeps working. Not mandatory — it depends on how important your project is.

Step 5: the orchestrator agent — the step you shouldn’t rush

Here, the video deliberately stops, because this is the most important step of the setup. The first agent created is the orchestrator: the one that will pilot all the others. And the classic mistake — the one I made — is to charge ahead: “first launch, the agent is alive” and boom, you create an empty, mute agent that will do whatever it wants, its own way.

The right method: open a separate conversation with an AI (Claude, ChatGPT, whatever) and tell it: “I’m creating my first Hermes orchestrator agent. Ask me every question needed to prepare this creation.” It will walk you through a series of questions about your project, your intentions, your needs — you answer as you wish, then you ask it: “Write me a complete and precise prompt to create this AI agent.”

You send this prompt to Hermes (via your desktop AI) and it creates the orchestrator shaped like you. Examples of my own choices: concise answers without fluff, tables whenever possible to organize information, a general theme the agent knows about. Yours might be sports coaching or something else — what matters is that it’s yours.

A painful lesson learned from creating my agents on an assembly line: as soon as the first one works, you get carried away and create tons of them. Result: the agents mix data between themselves, it’s very leaky. If you want several siloed projects, you’ll need Docker containers — independent bubbles on the VPS. But that’s another story: at the start, a single orchestrator.

Step 6: Discord — your daily interface with the agent

Hermes has its own apps (PC and mobile, in development), but I don’t use any of them: I switched to Discord from the start, after a Telegram experience I didn’t enjoy. It’s a choice, not an obligation — but Discord does the job perfectly, and above all it supports voice, which changes everything (step 7).

The steps to follow, always guided by your AI:

• Go to the Discord Developer Portal, create a new application, then a bot.
• Reset the bot’s token (yes, “reset” is the official term for the first display — it’s confusing but normal), copy it and paste it into your AI conversation so it configures Hermes.
• In the privileged intents, validate everything. In the bot’s permissions, give it admin. You’re creating this Discord server only for your agent — no matter if it has more rights than needed.
• Create a Discord server “for my friends and me”, right-click on its name → copy the server ID, paste the ID into your AI conversation, specify the channel (general).
• Your AI generates a bot invitation link: click it, validate, and your bot appears online in your channel.

The bonus bug that will spare you an hour of doubt: if you write to your bot and it doesn’t answer, that’s normal. By default, it doesn’t listen to the channel — it isn’t spying on you. Go back to your AI and tell it: “My agent doesn’t listen to my messages on Discord, set that up.” Thirty seconds later, it’s fixed.

Step 7: voice — your assistant in your pocket

This is the detail that transforms the experience. Ask your agent directly: “How do I set up voice between us in Discord?” It installs by itself a local transcription tool on the VPS — by its own admission in the video, Faster Whisper, local, zero API key, small fr model. It doesn’t burn tokens.

Concrete result: you’re in the subway, an idea crosses your mind, you open Discord, you hold the mic button, you talk. When you get home, it’s in place. Voice messages arrive transcribed automatically. You can also ask it to answer you by voice — for my part I only use it one way, I prefer reading its answers. Live voice isn’t supported, and honestly I don’t miss it.

And there, you understand what a permanent assistant really is: leaving it reminders (“every Monday at 10am, remind me to do this”), regular activities, ideas you don’t want to lose. The possibilities follow your activities, not the other way around.

What it actually costs

The real numbers of my setup, exactly as they are:

• VPS: €3.60/month (Contabo Storage VPS 10) — the agent quotes “€10/month” in the video to be safe, the truth is even lower.
• LLM: ~$15/month (GLM 3-month subscription) — or $20/month if you go with ChatGPT Pro, with image generation as a bonus.
• Installation time: budget half an hour following the guide — we announced 2 hours “just in case”, in reality it goes faster.

The total budget therefore comes to around €20/month for a permanent, personal assistant that lives on your machine. Compared to an employee, the math is quick. Compared to an all-in-one SaaS agent subscription, you mainly gain ownership: it’s your server, your agent, your data.

The questions I get asked the most

Do you need to know how to code to install Hermes?

No. That’s the whole point of the method: your desktop AI (Claude, Antigravity, Codex) drives the server over SSH on your behalf. You never open a terminal yourself — you ask, you copy-paste when it gives you a command, you validate.

Which VPS should you choose to get started?

The cheapest one, running Ubuntu, hosted in Europe. An entry-level VPS (e.g. 100 GB SSD, 8 GB RAM) is more than enough; I personally chose a storage plan at €3.60/month for the disk space. You can always upgrade later.

Which LLM should you plug into Hermes?

ChatGPT Pro if you’re starting out (the simplest, images included), Claude if you already have the subscription, a DeepSeek API key for something cheap on usage, or GLM 5.3 Flash like me — reliable and consistent in the background, which matters more than raw intelligence for an agent.

My Discord bot doesn’t answer, what should I do?

It’s the setup classic: by default the bot doesn’t listen to the channel. Just tell your AI: “my agent doesn’t listen to my messages on Discord, set that up” — the configuration takes a few moments.

Can you create several agents on the same VPS?

Yes, but not from the start. I made the mistake of creating agents on an assembly line and they mixed data between projects. One orchestrator first; if you then want to silo projects, we’ll talk about Docker in a future video.

The final word

The basic setup ends here: secured VPS, Hermes installed, LLM plugged in, orchestrator created, Discord connected, voice working. What comes next — the skills, the second brain, the memory, the habits I’ve built on top of this foundation — will be the subject of a future video.

Until then, if you want to see what a Hermes agent looks like in real conditions, I’ve documented how I run part of my business remotely with my Hermes agent on a VPS, and the AI Agents hub gathers everything I’ve tested on the topic. If the installation tempts you but the LLM budget holds you back, I’ve also tested DeepSeek open source to work on your own VPS.