How to Self-Host Hermes Agent on a VPS (2026 Setup)

Hermes Agent makes a lot of unusual claims for an open-source project: it remembers everything, it writes its own skills, it actually self-improves. After three weeks of running it on a Hetzner CCX13, the claims hold up. Here is the install path that actually works in 2026, with the gotchas the official docs glide over.

What You Need Before Starting

• A VPS with Ubuntu 24.04 or Debian 12 (Hermes also supports macOS and WSL2, this guide is Linux-only)
• SSH access with sudo or root
• 4 GB RAM minimum if you bring your own model, 16 GB if you self-host the LLM
• 80 GB disk, NVMe preferred
• A model endpoint (OpenAI API key, Anthropic API key, or a local Ollama install)

The VPS sizing depends on whether you self-host the model. See the best VPS for Hermes Agent guide for the realistic shortlist. For this tutorial I assume the bring-your-own-model setup, which is what 80% of users actually do.

Step 1: Provision the VPS

SSH into a fresh Hetzner CCX13 or equivalent and update the package index:

sudo apt update && sudo apt upgrade -y
sudo apt install -y curl git ufw

Set up a basic firewall before exposing anything to the internet:

sudo ufw allow ssh
sudo ufw allow 443/tcp
sudo ufw enable

Skip port 80 unless you plan to run a public web UI. The messaging gateways (Telegram, Slack, Discord) all use outbound connections and do not need inbound ports open.

Step 2: Run the Hermes Install Script

The official one-liner from the Nous Research repo:

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

This script does several things behind the scenes:

1. Creates a hermes system user
2. Clones the Hermes repo to /opt/hermes
3. Installs Python 3.11 and dependencies via a managed virtual environment
4. Sets up SQLite for the skill store
5. Registers hermes.service as a systemd unit
6. Prompts for your model endpoint configuration

The prompt for model endpoint asks for an OpenAI-compatible URL and an API key. For OpenAI direct: https://api.openai.com/v1 and your key. For Anthropic via a proxy: the proxy URL. For local Ollama: http://localhost:11434/v1.

Step 3: Verify the Install

Check the service is running:

sudo systemctl status hermes

You should see active (running). If not, the most common cause is the model endpoint config. Check the log:

sudo journalctl -u hermes -n 50

Common errors at this stage:

• 401 Unauthorized: API key wrong, edit /etc/hermes/config.toml
• Connection refused: model endpoint URL wrong, especially for local Ollama setups
• Permission denied on /opt/hermes/skills.db: SQLite file permission issue, fix with sudo chown hermes:hermes /opt/hermes/skills.db

Step 4: Add a Messaging Gateway

Telegram is the easiest first integration. Create a bot via @BotFather on Telegram and grab the token. Then:

sudo nano /etc/hermes/gateways/telegram.toml

Add:

[telegram]
token = "YOUR_BOT_TOKEN"
allowed_users = ["YOUR_TELEGRAM_USER_ID"]

The allowed_users list is important. Without it, anyone on Telegram who finds your bot can talk to Hermes and rack up your model API costs. Get your Telegram user ID by messaging @userinfobot.

Reload Hermes:

sudo systemctl restart hermes

Send a message to your bot. Hermes responds within a few seconds. The first message takes longer because the agent loads context.

Step 5: Test the GEPA Self-Improvement

Ask Hermes to do something non-trivial. A good first test: “Search GitHub for the trending Python repositories this week and summarize the top three.”

After Hermes completes the task, check the skill store:

ls /opt/hermes/skills/

You should see a new skill file generated. This is GEPA in action, Hermes wrote a reusable procedure for itself. The next time you ask a similar question, it uses the stored skill instead of reasoning from scratch. After 20 to 30 successful tasks, you will see measurable speedup on repeated work.

Step 6: Set Up Automatic Backups

The skill store is the irreplaceable part of your Hermes install. Back it up regularly:

sudo crontab -e

Add:

0 3 * * * tar -czf /backup/hermes-$(date +\%Y\%m\%d).tar.gz /opt/hermes/skills.db /opt/hermes/skills/ && find /backup -name "hermes-*.tar.gz" -mtime +14 -delete

This creates daily backups at 3 AM and keeps 14 days. Adjust the destination to wherever you store backups (S3, Backblaze, a separate disk).

What Comes Next

The first week of Hermes use is mostly about adding integrations (Slack, Discord, calendar, GitHub) and watching it build up skills. Around week three, the speedup from accumulated skills starts being noticeable. Around week six, Hermes is meaningfully faster at your recurring tasks than a fresh GPT-5 prompt would be.

Some integrations worth adding next:

• Slack gateway for team use
• Filesystem tool for document work
• Calendar integration for scheduling
• GitHub MCP for code-related queries

For the broader picture of where Hermes fits in the agent ecosystem, see the best VPS for Hermes Agent comparison. For other AI agent setups, check the SelfHostVPS guides.