A practical introduction

Hermes

An AI agent that stays awake on your own server and answers from your chat app. Plus the five ideas you need to understand it.

you, on Telegram Hermes, on a server a language model

Foundations

Five words, each built on the one before

AI

Software that does tasks we used to think needed a person.

LLM

One kind of AI: a model trained on text that writes text back.

Agent

An LLM that can take actions and keep going until a job is done.

Harness

The program around the model that gives it tools, memory and rules.Hermes is one of these

TUI

A harness you use by typing in a terminal window.

The next five slides take these one at a time.

Foundations · LLM

An LLM writes one small piece of text at a time

prompt
The capital of Tamil Nadu is Chennai,acoastalcityontheBayofBengal.

Each coloured piece is a token, roughly three quarters of a word. The model picks the most fitting next token, again and again.

Tokens are also the unit you pay in

Input$4

Output$20

Per million tokens, the rates quoted in our session for a top-tier model. What the model writes costs about five times what it reads.

Subscription: a flat monthly fee, like ChatGPT or Claude plans.
API key: pay only for the tokens you use.
OpenRouter: one wallet and one key for hundreds of models, some of them free.

Foundations · LLM

The context window is the model's working memory

The model remembers nothing between messages. Every turn, the whole conversation so far is sent again.

window full → conversation summarised, detail lost

0each stripe is one more turn272K tokens

It has a size

Common limits run from about 272K tokens to 1 million. The harness shows how full it is, for example 13.4%.

It fills up

At the limit, the harness squeezes the chat into a summary and starts a fresh window. Fine details go missing.

So match it to the job

Short questions fit anywhere. Long projects want the larger window, and a new session per topic keeps things sharp.

Foundations · Agent

An agent is an LLM working in a loop

repeat
until the job
is done

Think Act Observe

A chatbot answers once. An agent keeps going.

What we asked: “My Hermes app is not starting. Find the issue and fix it.”

Think

Locate the installation and try to start it.

Act

Runs the start command in the terminal.

Observe

The backend times out during startup.

Think

A half-finished update is forcing a full rebuild.

Act

Completes the update and relaunches.

Observe

Starts normally. Done, with no steps from us.

Foundations · Harness

The model is the brain. The harness gives it hands.

A model on its own

Knows a lot and reasons well
Can only produce text
Cannot open a file, run a command or browse

What the harness adds

Tools: actions it can take, such as the terminal, files and web search
Skills: saved instructions for a job it has learned to do
Memory and permissions, and it runs the loop

you say“Create a folder called reports.”The model decides the command. The harness runs mkdir reports.

you say“Organise this folder by client.”The model writes a small script for your rule. The harness runs it and reports back.

Hermes is not an LLM. It borrows its intelligence from whichever model you connect.

Foundations · TUI

A TUI is a harness that lives in your terminal

> add a holidays page to my site

● Read 3 files

● Wrote holidays/index.html

● Ran git push

✓ Page is live

↑6.2k ↓138 · $0.036 · context 13.4% · thinking: medium

The status line is real: tokens in, tokens out, cost so far, how full the context window is, and how hard the model thinks before answering.

Harness	From	Models
Claude Code	Anthropic	Claude only
Codex	OpenAI	GPT only
Pi, OpenCode	Open source	Any you bring
Hermes	Open source	Any you bring

There are 20 to 30 of these today and a new one most weeks. They differ in polish and features. The loop underneath is the same.

Why Hermes

The catch with a terminal agent: close the laptop and it stops

	Agent in your terminal	Agent on a server, in your chat app
Awake	Only while the laptop is open	All day, every day
You reach it	At your desk, in the right folder	From your phone, like texting a colleague
It starts work	When you type	When you message, or on a schedule
It works on	Your personal files	A machine of its own

The right-hand column is what Hermes is built for.

Hermes

Hermes is an open-source harness built to be a personal agent

Same loop as a coding TUI, with a desktop app, a web UI and a gateway to your messaging apps on top.

Lives in your chat apps

Telegram, Slack, WhatsApp, Matrix and more. Telegram is the easiest to set up.

Any model

Pick a provider, sign in with a subscription or paste an API key, and switch whenever you like.

Memory

Tell it your name, interests and how you work. It keeps that across conversations.

Skills

Teach a workflow once, such as “how I publish a page”, and ask for it by name afterwards.

Scheduled jobs

Recurring tasks run on their own and post results to a home channel you choose.

Yours

Your server, your configuration, your data. You always know which model sees what.

Hermes · How it works

How one message travels

You

Type in Telegram on your phone

Bot

A doorway. It passes messages along and has no intelligence.

Hermes agent

Runs the loop on your server

Tools and skills

Files, terminal, web search, GitHub

LLM provider

Decides what to say and what to do next

Bot = the doorway

Connects you to the agent. Nothing more.

Agent = the worker

Takes the request, uses tools, keeps going.

Model = the brain

Consulted at every step, even the simple ones.

Hermes · Where it runs

Give the agent a machine of its own

A VPS is a rented computer in a data centre. You will also hear “VM” and “instance”. They mean the same thing here.

Always on

The cloud provider keeps it powered and online. No dead battery, no closed lid.

A safe playground

The agent can create, change and delete freely. If it makes a mess, your own files are untouched and you rebuild the server.

Cheap enough

AWS, Google Cloud and Azure are priced for companies. Hetzner, Contabo and DigitalOcean suit individuals. Ours: 8 GB, in Mumbai, about ₹1,300 a month.

Pick a region near you for faster replies. Two shortcuts exist: Hermes offers a one-click cloud deploy, and it also runs on your own computer if that stays switched on.

Setup

The setup, in six steps

1

Rent a server

A small Linux VPS from any provider.

2

Log in

Connect from your laptop with SSH. It feels like a local terminal.

3

Install

Copy the one-line Linux installer from the Hermes site and paste it.

4

Connect a model

Choose a provider. We signed in with a ChatGPT subscription using a device code.

5

Pick the settings

Choose the model and how hard it thinks. We chose medium.

6

Connect Telegram

Scan a QR code, name your bot, press Save and restart.

On the Mac desktop app, step 6 took about two minutes: scan, name the bot, say hi.

Setup · What really happened

It did not go smoothly, and that is the useful part

What broke	Why	What fixed it
The installer stopped halfway	The Node.js download failed a verification check because the server lacked a system package.	We started Claude Code on the server and said “my setup stopped midway, check and complete it”. It did.
hermes: command not found	The terminal had not picked up the newly installed program.	Reloading the shell profile with one `source` command.
The Telegram QR page would not open	The setup page ran on the server's own localhost:9119. Our laptop browser could not reach it, and the firewall blocks that port.	The agent printed the QR in the terminal, but pairing timed out. Still open: allow the port, or enter a bot token by hand.

We used an AI agent to install an AI agent. When you are stuck, ask the agent.

In use

Talking to it feels like texting

Chow Chow bot

hi

Hey, how can I help?

Give me a fact about Salem in Tamil Nadu

typing…

Salem is a major city in western Tamil Nadu, famous for its Malgova mangoes.

Behind that one reply

Bot

Hands the message to Hermes.

Hermes

Opens a session and picks the skill it needs.

Model

Thinks at medium effort and writes the answer.

Bot

Delivers it to your phone.

The same conversation appears in the Hermes app, so you can see exactly what the agent did. From here it can also set reminders, read photos you send and run research while you do something else.

In use

What people use it for

Morning briefing

Reads the news sites you chose and sends one summary before you wake up.

Inbox triage

With read-only access, tells you which of 100 emails need you.

Research on demand

“Analyse these 15 stocks and reply in this format.” Sent from your phone.

Publish a page

Writes the HTML, pushes it to GitHub, and the site redeploys itself.

Posts from your work

Reads the day's commits and drafts a tech post when something is worth sharing.

Server watch

Keeps an eye on containers and logs, and investigates when something fails.

Code review

Reviews pull requests automatically as they open.

Home and reminders

“Turn off all the lights.” “Remind me at six.”

A team of agents

One turns an idea into a spec, one writes tickets, one orders them, one builds.

The pattern in all nine: a job that repeats, needs some judgement, and does not need you watching.

Before you start

Before you hand it the keys

	Hermes	A hosted assistant
Model	You choose	They choose
Your data	On your server	On theirs
Custom skills	Anything you can describe	What they offer
Setup	You do all of it	Sign up and go

Hermes pays off if you enjoy tinkering. If you do not, a hosted assistant is a fair choice.

Four habits that keep it safe

Read-only first. Let it view mail before it can send or delete.
Its own machine. Keep it away from files you cannot lose.
Closed ports. Open only what the setup needs.
Watch the meter. Tokens cost money, and loops use many.

Start small: one bot, one job, read-only.