My little brother is setting up his own agent right now — from scratch, on his own machine. This is what I handed him: a film that is a tutorial and an ad in the same breath, and the same the same steps on one A4.
Ten steps, in the order they are needed
On screen the commands are English, the voice is Swedish. It takes a computer, a terminal window and a bank card for five euros — nothing else.
- Install — one curl command in the terminal. macOS, Linux or WSL2; on Windows there is a PowerShell line instead.
- hermes setup — pick a model, pick a voice. It asks for the rest.
- DeepSeek — an account at platform.deepseek.com, five euros on top, one API key. The model is
deepseek-flash: a million tokens of context, and it takes image and text in the same request. - Context — a million tokens: about eight novels in one conversation, without losing the beginning.
- Vision — it reads images. A photo of a tulip in, “that is a tulip” out. Image and text in the same request.
- CAD — ask it to draw a gear. It writes the code, runs it, and leaves a finished model.
- Web search — the Brave key. An agent that cannot read the web guesses, and a guess is more expensive than a search.
- Start —
hermesin the terminal, orhermes desktopfor the whole app. - Telegram — create a bot with @BotFather, get your ID from @userinfobot, hand Hermes both. It does the rest.
- The rest you talk your way to — tools, skills, scheduled jobs. A terminal, memory and cron are there from the first minute.
The cheat sheet
hermes-kommandon.pdf — one A4 page, in Swedish, the same commands in order. Print it and keep it next to the keyboard.
Running the same thing inside an organisation — where permissions, logs and spend caps have to be decided before the first prompt — is on the services page.