Hermes Agent Crash Course — VPS Setup (EASY)

Overview

Alejandro AO walks through a complete, end-to-end setup of Hermes Agent as an always-on AI personal assistant running on a VPS. The tutorial covers provisioning a cheap Hetzner server (€5/month), using an AI coding agent (Pi/Codex/Claude Code) to install and secure the VPS automatically via a reusable skill, connecting Hermes to Telegram for mobile access, brain-dumping workflows so the agent can suggest automations, setting up cron jobs for recurring research tasks, and managing everything through the built-in Hermes dashboard. The key idea: let an AI agent set up another AI agent, rather than doing manual server administration.

1 What Is Hermes Agent?

▶ 1:45

Hermes Agent is a self-improving AI agent designed to be always-on. Unlike coding assistants that only work while you're at the computer, Hermes runs continuously on a server and is accessible from anywhere through messaging platforms.

• Multiple communication channels — Telegram, Discord, Slack, WhatsApp, and more. You talk to it like a personal assistant (PA), not a coding tool
• Always on — runs on a VPS or home server, so your computer can be off and the agent keeps working
• Rich integrations via skills — comes with ~85 skills built-in, including productivity integrations (Notion, PowerPoint, Shopify), finance tools, and more
• ~25 built-in toolsets — web search, browser automation, terminal, file operations, code execution, vision, text-to-speech, cron jobs, sub-agents, and cross-platform messaging

2 Architecture Overview

▶ 3:08

Alejandro provides a clear overview of how Hermes processes every message in a standard ReAct loop:

1. Message received — the agent receives your message from Telegram (or any platform)
2. System prompt built — for each thread, Hermes dynamically constructs a system prompt containing: available tools, loaded skills, current memory about you, project context, and platform context
3. LLM reasoning — the system prompt + message go to the language model, which either calls a tool or generates a response
4. Tool loop — if a tool is called, the result feeds back to the LLM, repeating until a final response is ready
5. Memory update — after responding, the agent updates its memory if you shared preferences, workflow details, or personal information
6. Future sessions benefit — that updated memory is injected into future system prompts, so the agent gets better over time

Conversations in Telegram stay in the same thread by default unless you explicitly ask Hermes to restart the thread.

3 Provisioning a VPS on Hetzner

▶ 4:37

Alejandro chooses Hetzner as the VPS provider, praising its developer experience and affordability. He also mentions OVH and DigitalOcean as alternatives.

Server specs

• Plan: €5/month (cost-optimized)
• RAM: 4 GB
• Storage: 40 GB
• Traffic: 20 TB
• Billing: per-hour with monthly cap — if you shut down mid-month, you only pay for the hours used
• OS: Ubuntu
• Networking: IPv4 + IPv6

SSH key setup

Before creating the server, generate an SSH key pair:

ssh-keygen -t ed25519 -C "hetzner-vps"

Save it in ~/.ssh/ with a descriptive name (e.g., hetzner_vps_key). Copy the public key content and paste it into the Hetzner SSH key field during server creation.

4 Cloud Config & Initial Security

▶ 7:40

During server creation, Alejandro uses Hetzner's Cloud Config (cloud-init) to pre-configure the server on first boot:

• User creation — creates a named user (e.g., "alejandro") with sudo group access and /bin/bash shell
• SSH authorized keys — injects the public key so the user can SSH in immediately
• Package updates — package_update: true runs apt update on first boot
• Baseline packages — installs curl, git, ufw, and fail2ban
• Firewall rules — runs ufw allow OpenSSH and ufw --force enable on install

After the server is created (takes a few seconds), connect via SSH:

ssh -i ~/.ssh/hetzner_vps_key alejandro@<SERVER_IP>

5 Agent-Driven VPS Setup with Pi + Codex

▶ 10:07

This is the core innovation of the tutorial: using an AI coding agent to install and configure the entire VPS instead of running commands manually. Alejandro installs Pi (a minimalist coding agent) on the fresh VPS and gives it a reusable VPS setup skill.

Installing Pi

1. Install Pi from pi.dev — it auto-installs Node.js as a dependency
2. Source ~/.bashrc after installation so the pi command is available
3. Log in with a ChatGPT subscription (or another provider) — Alejandro uses GPT-5.5 on low/no reasoning
4. Install the pi-web-access extension for web research capabilities

The VPS setup skill

Rather than typing individual commands, Alejandro copies the raw content of his hermes-vps-setup skill and pastes it directly into Pi with the instruction: "Help me set up this new VPS. Current user has sudo privileges."

The skill orchestrates the entire setup in phases:

• Phase 1 — Inspection: detects OS, current user, SSH status, and privileges
• Phase 2 — Baseline security: updates packages, installs tools (git, curl, tmux), configures UFW firewall (allows SSH, HTTP, HTTPS), enables fail2ban, disables root SSH, disables password SSH (after confirming key access works)
• Phase 3 — Runtime dependencies: installs Docker Compose, Node.js, FFmpeg, and Playwright browser dependencies
• Phase 4 — Agent user: creates a dedicated agentuser with a home directory containing .hermes/ for Hermes installation

6 VPS Security Hardening & Hermes Installation

▶ 13:30

During Phase 2, the agent performs critical security hardening with interactive confirmation:

• Firewall (UFW) — allows only SSH, HTTP, and HTTPS traffic. HTTP/HTTPS needed for the dashboard
• Fail2ban — protects against brute-force SSH attacks by banning repeat offenders
• Disable root SSH login — eliminates the most common attack vector
• Disable password SSH login — the agent asks for explicit confirmation first: "I found an SSH authorized key for Alejandro, so I should be safe to disable password SSH login. This reduces brute force risk, but if your SSH key does not work, it could lock you out. Please confirm."

Hermes installation

After security hardening, the agent installs Hermes under the agentuser account. It reports back:

• Hermes installed successfully at /home/agentuser/.hermes/hermes-agent
• Playwright Chromium installed for browser tools
• Hermes can update and modify itself because the CLI binary runs directly from the source code
• hermes doctor shows expected pre-setup warnings (missing Telegram/Discord config)

The next step requires an interactive terminal — the setup wizard can't run inside Pi, so you SSH in directly.

7 Hermes Setup Wizard — Configuration Deep Dive

▶ 19:10

Switch to the agentuser and run the interactive setup wizard:

sudo -u agentuser -i bash -c "cd ~/.hermes/hermes-agent && ./bin/hermes setup"

Select "Full setup" and configure each option:

Inference provider

Options include OpenRouter, Anthropic, OpenAI, and Hugging Face Inference Providers. Alejandro chooses Hugging Face:

• Create a HF access token at huggingface.co → account → new token
• Important: enable "Make calls to inference providers" permission
• Base URL: router.huggingface.co/v1
• Model: MiniMax M2.7 — affordable and very good quality
• Fallback provider: optional, skipped in this tutorial

Text-to-speech

Default is Edge TTS (free). Alejandro recommends ElevenLabs for premium, human-like voice quality if you have a subscription.

Terminal backend

Keep local since the VPS setup skill already containerized everything with the dedicated agent user.

Maximum iterations

Default is 60, Alejandro sets it to 90 — enables more complex workflows while preventing infinite loops that drain tokens.

Tool progress mode

Options: every tool call, new tools only, or none. Alejandro prefers "new" — if the agent calls web search 10 times, you only see the notification once.

Context compression

Hermes auto-compresses conversation when reaching a percentage of the context window:

• MiniMax M2.7 has ~200K tokens → set to 0.90 (compress at 90%)
• For 1M-token models → set to 0.5 or lower
• Smaller context windows need higher thresholds to avoid premature compression

Session reset mode

• Inactivity timeout: 1440 minutes (24 hours — daily reset recommended)
• Daily reset hour: 4 AM local time ("I'm rarely working at 4 AM")
• The agent retains memory of previous sessions, so resetting just clears the active conversation thread

Tools selection

Alejandro enables: web search, browser automation, terminal, file operations, code execution, text-to-speech, skills, memory, session search, clarifying questions, task delegation (sub-agents), cron jobs, and cross-platform messaging. He skips vision/image generation.

Search provider

Options: self-hosted Firecrawl (via Docker — more complex), Exa, or cloud Firecrawl. Alejandro uses a Firecrawl API key for simplicity.

8 Telegram Integration

▶ 26:03

With Hermes configured, the next step is connecting it to Telegram as the primary communication channel.

Gateway setup

Switch to agentuser and run:

sudo -u agentuser -i bash -c "cd ~/.hermes/hermes-agent && ./bin/hermes setup gateway"

Select Telegram with spacebar, then hit enter.

Creating a Telegram bot

1. Open Telegram and search for @BotFather
2. Send /newbot
3. Give it a display name (e.g., "Bobby")
4. Create a unique username ending in bot
5. Copy the bot token and paste it into the Hermes setup

Restricting access

Without restrictions, anyone could message your bot and access your personal AI agent. To lock it down:

1. Search for @userinfobot in Telegram
2. Send /start to get your personal Telegram user ID
3. Paste the ID as the allowed user in the Hermes gateway config
4. Set it as the home channel (your primary conversation destination)

Running as a system service

When asked, select "system service" — this ensures the Telegram gateway starts automatically on boot and runs in the background permanently. After setup, go to your bot in Telegram, hit /start, and you're connected.

9 Brain Dump Workflow

▶ 30:53

Alejandro demonstrates a powerful onboarding pattern: brain-dumping your entire workflow into Hermes and asking it what it can help with.

His prompt: "I'm going to tell you a few things about my workflows, so can you tell me where you can help me out and how we can work together?"

After describing his daily workflow, Hermes suggested help with:

• Video production pipeline — research, scripting, metadata generation
• Quick side project scaffolding — "If I have an idea for a web application, just have Hermes build it"
• General research — on-demand deep dives into any topic
• Cron job automations — recurring scheduled tasks (shown next)

10 Cron Jobs — Automated Research

▶ 32:25

Alejandro creates a recurring cron job directly through Telegram chat — no terminal needed:

His prompt: "Create a cron automation that runs every three days. On each run, research developments for the past 3-5 days in AI research and product launches, open source AI models, etc. Identify tutorial ideas and for each run produce a concise summary, 5-10 tutorial ideas. Prioritize ideas that are hands-on, reproducible, open-source friendly, and likely to perform well with the Hugging Face community."

What Hermes did

• Created the cron job with a unique ID
• Scheduled to run every 3 days
• Output delivery: Telegram DM
• Offered to run it once immediately to preview the output

The first run took about 4 minutes, producing tutorial ideas with sources split across multiple Telegram messages. From now on, Alejandro receives this research brief every 3 days automatically.

11 Dashboard Setup & Management

▶ 34:14

The final piece is the Hermes web dashboard — a management UI accessible at the VPS's IP address.

Enabling the dashboard

Alejandro goes back to his Pi session (which has full context of the VPS setup) and asks it to set up the dashboard service. Since the VPS setup skill already includes dashboard instructions, Pi knows exactly how to configure it:

• Sets up the dashboard as a system service
• Secures it with HTTP basic authentication (username + password)
• Makes it accessible at https://<SERVER_IP>

Dashboard features

Once logged in, the dashboard provides:

• Connected platforms — shows Telegram (and any other platforms configured)
• Sessions — active sessions (Telegram conversations + cron job runs)
• Models — current main model (MiniMax M2.7 via Hugging Face)
• Logs — real-time agent logs for debugging
• Cron jobs — view, trigger, pause, or delete scheduled automations. You can also create cron jobs directly from the dashboard UI
• Skills — browse all 85 available skills
• Toolsets — 26 active toolsets with their status
• Plugins & profiles — manage extensions and configurations
• Configuration — model settings, context length (auto-detected), all configuration keys
• API keys — view which keys are configured
• Documentation — embedded iframe of the Hermes docs
• Kanban — experimental multi-task orchestration for sub-agents (teased for a follow-up video)

Alejandro's minimalism note

He notes that Hermes ships with ~85 skills and ~25 toolsets active from the start — more than he personally prefers. As a fan of Pi (a minimalist agent), he'd prefer starting with almost no tools and gradually adding ones he needs. But he acknowledges this "batteries included" approach works well for most users.

🎯 Key Takeaways

🔗 Resources & Links

• Written Tutorial — Alejandro's full written guide for Hermes Agent VPS setup
• VPS Setup Skill (SKILL.md) — the reusable skill that automates VPS provisioning, security hardening, and Hermes installation
• Hermes Agent Documentation — official docs for setup, configuration, skills, and integrations

⏱ Timestamp Index