Install Hermes Agent on a VPS: your 24/7 personal AI assistant

For months, I used OpenClaw as my personal assistant connected to Telegram, WhatsApp and Discord. Great to get started, but updates became painful with every release. I needed something more stable.

I switched to Hermes Agent (Nous Research, open source). And I don't regret it for a second.

In this article, I'm sharing everything I learned installing Hermes on a VPS to run 24/7, accessible from any phone — without depending on a third-party SaaS. Including the gotchas, the security trade-offs, and the setup I recommend.

You can have a personal AI agent running 24/7 on a VPS, accessible from your phone, with memory, skills, crons and GitHub backup.

Hermes is not Claude Code

Before diving into the install, let's clarify the positioning. That's probably your first question.

The one-liner that sums it up:

Claude Code is your copilot when you're driving. Hermes is the assistant who stays at the office while you're away.

Hermes is not a direct replacement for Claude Code. It's a personal always-on agent, plugged into channels like Telegram or WhatsApp, able to launch scheduled tasks and create its own skills on the fly.

Why a VPS?

A VPS is just a computer in the cloud. The point is simple: your agent no longer depends on your Mac or PC. When you close your laptop, the agent keeps running.

Concretely, this lets you:

• Talk to your agent from your phone, anywhere
• Run crons even when you sleep (monitoring, briefing, audits)
• Have an assistant answering in 2 seconds, not 10 minutes after your computer boots
• Isolate, monitor, back up cleanly

A VPS turns your agent into a service. Not just a chat window.

The 5 fundamentals of Hermes

Hermes is built on 5 building blocks that differentiate it from a plain LLM chat:

1. Memory — persistent memory across sessions
2. Skills — reusable capabilities (.md files the agent can create itself)
3. Soul — personality and tone
4. Sessions — search through past conversations
5. Crons — native scheduled tasks

The real "aha": your agent creates its own skills and runs tasks while you sleep. No longer a chat — an async collaborator.

3 installation modes

Before copy-pasting commands, pick your mode:

My pick: direct install with a dedicated non-sudo user. Consistent if you already have other services on the VPS (OpenClaw, n8n, personal APIs).

Step-by-step install (the setup I recommend)

1. Prepare the VPS

# as root, on Ubuntu 24.04 LTS
sudo apt update
sudo apt install -y curl git ripgrep ffmpeg build-essential

# create a dedicated non-sudo user
sudo adduser hermes
# NO usermod -aG sudo (strict segmentation)
# NO usermod -aG docker (root equivalent)

Why no sudo? An AI agent running under a user with sudo equals root through its terminal backend. If you share your VPS with other services, segment strictly.

2. Install Hermes

# as hermes
sudo su - hermes
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash

The installer will check and install:

• uv (which installs Python 3.11)
• Git, Node.js, ripgrep, ffmpeg

Hermes installs into /home/hermes/.hermes/hermes-agent. Remember this path for backups.

3. Configure the LLM provider

hermes setup

You can pick from: Anthropic, OpenAI, OpenRouter, MiniMax, Qwen, DeepSeek, Hugging Face, Bedrock, GitHub Copilot, custom endpoint, etc.

⚠️ Headless VPS gotcha: if you use Codex OAuth (ChatGPT), you must first enable in ChatGPT:

Settings → Secure Connection → "Enable device code authorization for Codex"

Without this toggle, you'll spin for 10 minutes.

4. Connect Telegram (or WhatsApp)

hermes gateway setup

For Telegram:

1. Create a bot via BotFather
2. Grab the bot token
3. Get your Telegram user ID (via @userinfobot)
4. Allowlist your user ID
5. Start the gateway
6. Test with a simple "hello"

⚠️ Allowlist is mandatory: a public bot without controls is a bad idea. The goal isn't for the whole internet to talk to your agent.

5. Working directory

This is the trap many miss. By default, the wizard suggests . (cwd at setup time). Don't leave .:

Gateway working directory [.]: /home/hermes/workspace

Otherwise, if systemd starts Hermes from /, your working dir becomes /. Catastrophic for crons writing in the wrong place.

6. Check it's running

hermes              # CLI chat
hermes --tui        # TUI interface
hermes doctor       # full diagnostics

If the basic chat doesn't work, don't add complexity. Debug first, extend later.

Security: the real trade-offs

This is the part most tutorials skip. An autonomous agent with terminal + API keys + Internet is not trivial.

Rules I follow systematically

• Dedicated API keys per agent (never the same key for all)
• Budget limits when the provider allows
• Minimum necessary access (zero blanket sudo)
• No full personal mailbox at start
• No prod without human validation
• Separate containers for agents with different roles

The "sudo enabled during the wizard" trap

The wizard asks "Enable sudo support? [y/N]". If you answer y, Hermes stores a password to run sudo apt install for you. Convenient, but breaks segmentation.

My call: start without sudo. If a need emerges (e.g. Hermes wants to install a system package), add a whitelisted sudoers rule — not blanket sudo.

# granular rule, not blanket
hermes ALL=(root) NOPASSWD: /usr/bin/apt, /bin/systemctl restart hermes-*

GitHub backup

Hermes will accumulate value: memory, skills, preferences, automations.

If your VPS dies and you have no backup, you lose your assistant.

The right pattern:

• Private GitHub repo for the agent state
• Strict .gitignore (never .env, tokens, sensitive logs)
• Regular commits (memory, soul, user, skills, docs)
• Daily backup cron

An agent's memory is an asset. Don't let it die on a VPS without a backup.

First useful cron

This is where Hermes really differs from an AI chat. A few concrete ideas:

• "Every morning, give me a 10-line AI/dev briefing"
• "For 12h after a YouTube video drops, watch comments and draft replies"
• "Every night, back up your skills and memory to GitHub"
• "Every Monday, run a basic VPS security audit"
• "Every hour, check if my SaaS has errors and send a summary on WhatsApp"

You're no longer just asking it to answer. You're asking it to come back on its own at the right time.

The 4 classic gotchas

Summary of everything I broke during the install:

These four mistakes are 80% of the time lost. Now you know.

Going further

I made a full video that shows the entire setup live, with 1h of free training: step-by-step install, root access, sudoers, exposed ports, Tailscale, OpenWebUI behind VPN, custom skills demo.

👉 Watch the full Hermes VPS video

And if you want to follow this kind of AI workflows for developers every week, subscribe to The Agentic Dev — my newsletter on AI agents and AI-assisted dev.