Hermes Agent v0.20: Complete Setup Guide 2026

Running your own J.A.R.V.I.S.-style AI agent on a €6 Hetzner VPS. Here’s every step to set up Hermes Agent v0.20 by Nous Research — from zero to a fully operational agent with Telegram, Voice Mode, and custom skills.

---

What is Hermes Agent?

Hermes Agent is an open-source, self-hosted AI agent platform by Nous Research. It’s not just a chatbot — it’s a complete agent framework with:

• Full tool access — file system, terminal, browsers, APIs, Docker
• Multi-platform delivery — WebUI, Telegram, Discord, WhatsApp
• Real-time Voice Mode — ElevenLabs TTS + Whisper STT
• Skill system — reusable procedural memory for recurring tasks
• Multi-agent orchestration — spawn sub-agents for parallel work
• Cron jobs — scheduled autonomous task execution

I’ve been running Hermes since v0.12 on my Hetzner infrastructure. Here’s the production-tested setup.

---

Prerequisites

Minimum Hardware

Tier	vCPU	RAM	Storage	Cost	Use Case
Test	2	4 GB	40 GB	€4-6/mo	Try it out, light tasks
Standard	4	8 GB	80 GB	€12-15/mo	Daily driver, n8n companion
Production	8	16 GB	160 GB	€30-40/mo	Multi-agent, voice mode

Software Prerequisites

• Ubuntu 22.04 or 24.04 (I use 24.04)
• Docker & Docker Compose (optional, for container deployment)
• Python 3.11+
• A domain with DNS pointing to your server (optional, for HTTPS WebUI)

---

Method 1: One-Line Install (Recommended)

Nous Research ships a bootstrap script that handles everything:

curl -fsSL https://get.hermes-agent.com | bash

This installs:

• Hermes Agent core (~/.hermes/)
• WebUI (port 8787)
• Default configuration with Claude Sonnet via OpenRouter

⚠️ My tip: The bootstrap script works perfectly on fresh Ubuntu 24.04. On existing servers with conflicting Python environments, use Method 2 instead.

---

Method 2: Docker Deployment

For production with n8n on the same server:

# docker-compose.hermes.yml
version: '3.8'
services:
  hermes:
    image: ghcr.io/nousresearch/hermes-agent:v0.20
    container_name: hermes-agent
    restart: unless-stopped
    ports:
      - "8787:8787"  # WebUI
      - "8192:8192"  # Box Terminal
    volumes:
      - ~/.hermes:/root/.hermes
      - ~/hermes-workspace:/root/hermes-workspace
      - /var/run/docker.sock:/var/run/docker.sock  # For Docker tools
    environment:
      - OPENROUTER_API_KEY=${OPENROUTER_API_KEY}
      - HERMES_DEFAULT_MODEL=openrouter/anthropic/claude-sonnet
    networks:
      - hermes-net

networks:
  hermes-net:
    external: true

# Start
docker compose -f docker-compose.hermes.yml up -d

# Check health
curl http://localhost:8787/health

---

Telegram Integration

This is the killer feature — Hermes responds to @mentions in Telegram:

# In ~/.hermes/config.yaml
platforms:
  telegram:
    bot_token: "8877787386:YOUR_TOKEN"
    allowed_chats: ""
    reactions: false

1. Create a bot via @BotFather on Telegram
2. Copy the token to ~/.hermes/bot_token (chmod 600)
3. Add the bot to your group chat
4. Restart Hermes: hermes gateway restart

Now you can work from your phone, anywhere in the world. I use this daily.

---

Voice Mode

Hermes v0.20 ships with a full voice pipeline:

# In ~/.hermes/config.yaml
voice:
  auto_tts: true
  provider: elevenlabs
  stt_provider: faster-whisper
  voice_id: Adam
  voice_mode_button: true

My production setup:

• STT: faster-whisper base (local, no API costs)
• TTS: ElevenLabs Adam (eleven_multilingual_v2)
• Streaming TTS + Barge-In: Active (interrupt the agent mid-sentence)

The WebUI has a dictation button — click, speak, auto-send. On Telegram, TTS replies play as native voice messages.

---

Skill System

Skills are Hermes’s procedural memory — reusable workflows you write once:

# Create a skill
hermes skill create server-healthcheck

# Edit it
vim ~/.hermes/skills/server-healthcheck/SKILL.md

Example skill structure:

---
description: "Use when checking server health. Runs diagnostics and reports."
---

## Steps
1. Check CPU, RAM, disk with `htop` and `df -h`
2. Check all Docker containers with `docker ps -a`
3. Report any stopped containers or high resource usage
4. Send summary to Telegram

I have 30+ skills for everything from n8n deployment to Zoho CRM health checks. They’re the difference between an AI assistant and an AI agent.

---

My Production Stack

┌─────────────────────────────────────────────────────┐
│                   Hetzner CX41 (8 vCPU, 16 GB)        │
├─────────────────────────────────────────────────────┤
│  ┌──────────┐  ┌──────────┐  ┌────────────────────┐ │
│  │  Hermes  │  │   n8n    │  │  Caddy Reverse     │ │
│  │ v0.20    │  │ v2.35.4  │  │  Proxy (443)       │ │
│  │ :8787    │  │ :5678    │  │  mbb.mbbserver.com │ │
│  └──────────┘  └──────────┘  └────────────────────┘ │
│  ┌──────────┐  ┌──────────┐  ┌────────────────────┐ │
│  │ GEX131   │  │  Postgres │  │  30+ Hermes        │ │
│  │ LLM Host │  │  17       │  │  Cron Jobs         │ │
│  │ :8000    │  │  :5432    │  │  (24/7 ops)        │ │
│  └──────────┘  └──────────┘  └────────────────────┘ │
└─────────────────────────────────────────────────────┘

• LLM Routing: Qwen 3.8 for German content, DeepSeek V4 for automation, GLM 5.3 for reasoning — all via Hermes skill routing
• Caddy: TLS termination, reverse proxy
• Tailscale: Secure remote access
• Telegram: Mobile interface, cron notifications, watchdog alerts

---

First 5 Minutes With Hermes

After installation, try this:

You: @brain status
Hermes: [Checks all systems, Docker containers, disk usage, recent errors]
        All systems nominal. 14 Docker containers running, 67% disk on /var.

You: @brain aktiviere Dev
Hermes: Dev ist aktiv. Was können wir angehen?

You: Check n8n logs for errors since this morning
Hermes: [Reads Docker logs, filters for ERROR, summarizes]
        3 errors found — all from executeCommand deprecation warning.
        No critical issues. Full log at /tmp/n8n-errors-2026-09-03.txt

---

Production Lessons

After running Hermes in production since v0.12:

1. Always use --restart unless-stopped for Docker containers. Server reboots from unattended-upgrades will kill containers without it.
2. Set up a healthcheck cron that verifies WebUI → Gateway → Telegram connectivity. Mail alerts on failure.
3. Use Hermes’s built-in sub-agent delegation for long-running tasks — don’t block the main agent.
4. Voice Mode + Tools don’t mix perfectly yet (v0.20). Use WebUI for complex tool work, Telegram for quick commands.
5. The skill system compounds. Each skill saves 5-15 minutes every time it’s reused. After 30 skills, you’ve saved dozens of hours.

---

Why Self-Host Instead of ChatGPT/Claude?

Feature	Hermes Agent	ChatGPT	Claude.ai
Runs your code	✅ Terminal, Docker, Python	❌ Sandbox only	❌ No execution
File system access	✅ Full read/write	❌ Uploads only	❌ Uploads only
Scheduled tasks	✅ Cron jobs	❌	❌
Multi-platform	✅ Telegram, Discord, WA	❌ App only	❌ Web only
Privacy	✅ All local	❌ Cloud	❌ Cloud
Cost	Server + API	$20/mo	$20/mo
Model choice	✅ Any (OpenRouter)	❌ GPT only	❌ Claude only

---

Next Steps

• Qwen 3.8 Flash-Next Installation Guide — Run a state-of-the-art local LLM
• GLM vs DeepSeek vs Qwen Comparison — Which model for which task
• KI-Integration for Business — Enterprise agent infrastructure

---

Steffen Hartmann is a Hermes AI Agent Expert and Automation Architect at MadeByBrain. He has been running Hermes in production since v0.12 and advises companies on local AI agent infrastructure.