How to Install Hermes Agent on a VPS (and Give It a Personality)

You want an AI agent that runs 24/7, remembers things across conversations, and doesn’t cost $200/month in API bills.

The answer is Hermes Agent on a VPS. Here’s how to do it without the headaches I ran into.

Why a VPS and not your laptop

Your laptop sleeps. Your laptop goes offline. Your laptop can’t receive a Telegram message at 3am when you’re asleep.

A VPS is just a computer in a data center that stays on. You rent it for about $5-10 a month. Hetzner, DigitalOcean, whatever. It has an IP address, you SSH into it, and you install things.

The important bit: you control it. No platform lock-in, no “we changed our pricing,” no one reading your conversations.

What you need

• A VPS with at least 2GB RAM (I use Hetzner, 4GB is better)
• Ubuntu 22.04 or 24.04
• An API key from a model provider (I use OpenCode Go)
• About 30 minutes

Step 1: Get a VPS

If you’re new to this, Hetzner Cloud is the sweet spot. Their CX22 (2 vCPU, 4GB RAM) is about €4/month. Create an account, spin up an Ubuntu instance, and you’ll get a root password by email.

SSH in:

ssh root@your-server-ip

First thing: update everything and create a non-root user. Running everything as root is asking for trouble.

apt update && apt upgrade -y
adduser deploy
usermod -aG sudo deploy

Now switch to that user and do the rest from there.

Step 2: Install Hermes

One command:

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

That’s it. The installer pulls in Python, Node.js, and everything else it needs. Takes about two minutes.

After it finishes, reload your shell and run the setup wizard:

source ~/.bashrc
hermes setup

Step 3: Pick your model provider

Hermes will ask which provider you want. The options: OpenAI, Anthropic, OpenRouter, OpenCode, Nous Portal, or a custom endpoint.

I use OpenCode Go. Here’s why.

OpenCode Go is a provider optimized for coding agents. The models are cheap and fast. For most agent tasks — file operations, research, writing, scheduling — you don’t need a $20/M-token model. You need something that responds in under a second and costs pennies.

My config uses deepseek-v4-pro through OpenCode Go. For the 5% of tasks that need real reasoning, Hermes can escalate to a stronger model automatically.

A few tips on model selection:

• Cheap model for 95% of work. OpenCode Go, DeepSeek, or Qwen. These handle file ops, research, and writing just fine.
• Expensive model on standby. Set up a fallback in config.yaml for when the agent hits something complex. I use o4-mini through OpenAI as my delegation model.
• Don’t overpay for cron jobs. Your daily briefings and git syncs don’t need Claude. Set those to run on the cheap model. I went from $130 to $10 per 5 days for always-on agent usage just by routing tasks to the right model tier.

Here’s what that looks like in config.yaml:

model:
  default: ""
  provider: opencode-go

fallback_model:
  provider: "openai"
  model: "gpt-4o"

delegation:
  provider: "openai"
  model: "o4-mini"

Step 4: Connect Telegram (or whatever you use)

Hermes talks to messaging platforms through a gateway. Telegram is the simplest to set up.

hermes gateway setup

Pick Telegram, create a bot through BotFather (Hermes walks you through it), paste the token. That’s it. Now you can text your agent from your phone.

One security tip: add your Telegram user ID to the allowlist in config.yaml. Otherwise anyone who discovers your bot can talk to your agent.

gateway:
  telegram:
    allowed_user_ids:
      - YOUR_TELEGRAM_ID

Step 5: Give it a SOUL

This is where most people stop. They have a working agent but it sounds like a generic chatbot. The SOUL file is what makes Hermes feel like your assistant, not just any assistant.

The SOUL.md lives at ~/.hermes/SOUL.md and gets injected at the start of every conversation. It defines:

• Who you are (so the agent knows its context)
• How it should talk (voice and tone)
• What it should and shouldn’t do
• How it should handle files and research

The anatomy of a good SOUL

Here’s the structure I use, adapted from actual consulting sessions:

Section 1: Who you are. Basic facts. Name, role, location, what you’re working on. The agent uses this to ground its responses in your reality rather than generic assumptions.

Section 2: Voice and tone. The single most underrated section. Without it, the agent defaults to helpful-corporate-speak. My rules:

• Be concise. No fluff
• Never use em dashes
• Write like you talk, not like a press release
• If it sounds like marketing copy, delete it

One rule that changed everything for me: “presence over persuasion.” The distinction between writing at someone (trying to convince) and writing near someone (stating what’s true). A good agent doesn’t sell you on things. It just says what it sees.

Section 3: Operating modes. The agent should switch behavior depending on the task. Mine has:

• Default mode: quick answers, file ops, get it done
• Caveman words mode: explain complex things simply, zero jargon
• Deep analysis mode: cross-reference everything, go hard
• Therapist-lite mode: when I bring up psychological patterns, analyze them properly

Section 4: Tools and rules. What it can do, what it can’t. Where files live. The git protocol (pull before read, commit after write). Research expectations.

Section 5: Push-back rules. This one matters. You want an agent that challenges you, not one that agrees with everything. My rules:

• When I undersell myself, remind me of my credentials
• When copy sounds like marketing, call it out
• When I compare myself to established competitors, remind me I’m at Day 1 and they’re at Year 5
• When I’m over-iterating instead of shipping, push me to ship

A real SOUL excerpt

From the section on how the agent should handle my work:

## How to Work

### Voice & Tone

The single most important rule: presence over persuasion. Write like
Marius talks, not like you're selling something.

Rules:
- Be concise. No fluff. Get to the point.
- Never use em dashes in any text.
- Lowercase energy even in formal contexts. No warmup.
- No moral, no lesson. State observations and let them land.
- Plain words over jargon. "Build" and "use," not "leverage."
- If it sounds like marketing copy, delete it.

See the pattern? Specific. Actionable. Tells the agent exactly what to avoid, not just what to do.

Step 6: The USER profile (shorter but critical)

Alongside the SOUL, Hermes keeps a compact user profile. This is facts about you, not instructions for the agent.

What to include:

• Preferences (avoid subscriptions, prefer self-hosted, default to open-source)
• Tech stack (what languages, what hosting, what tools)
• Working style (collaborative, ship-first, hates wasted time)
• Communication (language, directness, what frustrates you)

Keep it under 1,500 characters. The agent reads this every turn alongside the SOUL, so every character costs tokens.

Step 7: Run it as a service

You want Hermes running even when you’re not SSH’d in. Create a systemd service:

hermes gateway install
hermes gateway start

This keeps the gateway alive, reconnects if something crashes, and starts automatically on reboot.

To check if it’s running:

hermes gateway status

What I’d do differently now

After setting up Hermes for myself and for clients, a few things I’d change if I started fresh:

Start with the SOUL, not the tools. The first hour should go into defining voice, rules, and context. Tools can be added later. A well-configured SOUL makes every subsequent interaction better.

Use the cheap model first. New users tend to pick Claude or GPT-4 for everything and get a $200 bill week one. Start with OpenCode Go or DeepSeek. Only escalate when the agent actually fails a task.

Allowlist Telegram immediately. I’ve seen people skip this and then wonder why their agent is responding to strangers. It takes 30 seconds. Do it.

Set up cron jobs for the boring stuff. Daily git sync, morning briefings, weekly summaries. If you have to remember to tell your agent to do something, the setup is incomplete.

What’s next

Once Hermes is running, the real work starts: skills, cron jobs, the Obsidian vault. But this gets you from zero to a working, personality-driven agent in under an hour.

If you want help setting this up for your business, I do this as a consultancy. We get on a call, I watch you type the commands (you learn faster that way), and you walk away with a running system. Let’s talk.