Source: raw/Hermes_Agent_-_Full_Setup_Guide_Beginner_From_Scratch.md (auto-caption transcript)

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=CcsgednSRfE · Duration: 19:44 · Uploaded: 2026-10-09 · Sponsor: Hostinger (one-click Hermes VPS; promo code for 10% off). He also links a free prompt doc and his newsletter.

A no-code walkthrough that takes Hermes from nothing to an agent that messages you first. Sharbel frames it as hiring an employee: a place to live, a brain, a phone, a job and a schedule. The useful parts are his three model tiers, the Telegram-group-with-topics layout, and a morning-priority cron that he also runs himself.

Key Takeaways

  • What makes Hermes different from a chatbot, in his words: it runs on its own (schedules, messages you first), it remembers across devices, and it saves solved tasks as skills.
  • Where it lives. Free: a spare always-on machine (his is a Mac Mini). Paid: a cloud server, either the one-click host on the Hermes site (via Nous Portal) or Hostinger, which deploys Hermes in one click. He suggests about 8 GB of RAM (Hostinger’s KVM 2 plan) and keeping the username hermes.
  • Two safety rules before you start. Give the agent its own machine, not one holding your personal files, and keep approval gates on, because a web page it reads can attempt prompt injection.
  • Three model tiers; match the model to the job.
    • Top (the boss): GPT-6 Astra or 6.1 Astra, Claude Opus 5.5. Hard decisions, coding, anything where a mistake costs money.
    • Middle (the workhorse): GPT-5.6 Sol or Luna, Claude Sonnet 5.5 or Haiku. Planning and everyday jobs.
    • Bottom (grunt work): cheap or free models for sorting, tagging, summarising, formatting. “A model being free doesn’t make it smart”; test before handing it thinking.
  • Ways to connect a brain. OpenRouter models tagged free (rate-limited), a small local model through Ollama, pay-as-you-go on OpenRouter (“$5 on a cheap model lasts weeks”), or your existing ChatGPT plan through the Codex subscription login in the dashboard’s Keys page. His pay-as-you-go pick: DeepSeek V4.1 Flash, with Kimi and GLM as alternatives.
  • Use a Telegram group with topics, not a DM. One agent, one memory, separate rooms: general, an email feed (leads and drafts), alerts (only when something breaks or a watched thing changes), content. Pairing is a QR scan in the Channels page; that adds you to the allow list.
  • Corrections become permanent. Say “That’s too long. Keep it short. No bullet points” and tell it to remember. When it does something well, tell it to save that as a skill if it did not already.
  • Money and safety. Server: a few dollars a month, or free on your own machine. Set a spending limit at your model provider. Put approval gates on anything that spends money, sends to a client or deletes.
  • Start with one agent and one real job. Bot teams and separate work/personal profiles come later.

Try It

  1. Install the desktop app from the Hermes site, or run hermes dashboard in a terminal to get the same dashboard on a server.
  2. In Keys, connect a model (your Codex subscription if you have ChatGPT; otherwise a free OpenRouter model or DeepSeek V4.1 Flash), then send “hello”.
  3. Pair Telegram from Channels, set it as the home channel, then move Hermes into a group with topics.
  4. Send his morning-priority prompt (from the video): “Every morning at 7:00 a.m., message me and ask my number one priority for the day. When I reply, start working on it right away and get as far as you safely can before I sit down. Don’t spend money or send anything externally without asking me first.”
  5. Try a watcher he built live: ask it to watch your niche hourly, build itself a skill for the check, track what it has already shown you, and only message you when something is genuinely new.

Implementation

Tool/Service: Hermes Agent (free, open source); hosting on your own machine, Nous Portal’s one-click cloud, or Hostinger’s one-click VPS. Setup: Dashboard pages: Sessions (conversations), Models, Cron (schedules), Skills, Keys, Channels. Cost: Hermes itself is free. Hostinger plan prices were shown on screen, not spoken; the 12-month option adds 1,000 free Oxylabs web-scraping credits, per the video. Model spend depends on the tier you pick. Integration notes: He reports an example of the morning job turning a voice-note reply into an email watcher that sorts leads into groups with Jev; see Jev.

Open Questions

  • Hostinger’s KVM 2 price and the Nous Portal free-plan limits are not given in the transcript.
  • “GPT 5.6 Soul” in the captions is read here as GPT-5.6 Sol; check against GPT-6 Astra, Sol and Luna.