The video describes Hermes Agent as an open-source AI agent from Nous Research. The author claims it can sit on a small cloud server and run all day, act on a schedule, message you through Telegram, and turn what it has done into reusable skills. For a beginner, three things matter at the start: how to set it up, what it will really cost each month, and which of your files it can reach. This article is based on the Arabic subtitle track of a beginner tutorial from the YouTube channel Metics Media. Every step, price, and version below is a translation from that subtitle track, not checked against the original video, so please rely on official documentation for anything you act on.
Picture this at four in the morning: your laptop is closed, your phone buzzes, and it is a message from the agent. That is the scenario the author demonstrates. The agent checks a topic every few hours and only notifies him when something new shows up; in his demo the result was eight new videos. This passage is an illustration, not a result you are guaranteed to get by following along. It is also exactly where the author says this differs from an ordinary chatbot: it does not wait for you to speak first.
But pause for a moment. Something that browses the web, writes files, and runs programs for you has a very different cost and risk profile from a chat window.
How Hermes Agent Differs From a Regular Chatbot
The author lists three differences, as rendered by the subtitle track.
The first is continuous operation: it can work on its own schedule and start a conversation itself instead of waiting for your question. The second is memory: preferences you tell it are kept across conversations and across devices. The third is self-improvement: after finishing a complex task, it records the steps as a skill and reuses it later.
The author also says the software itself is free and under the MIT license, and that you pay for only two things: model usage, and the machine it runs on. That is the author's claim. Please check the license text yourself.
Setup Flow: Where to Put It and How to Install
Decide where it will live first. The author gives beginners two options. The first is the official desktop app, installed directly on your computer. It is the fastest way to try it, but once the laptop goes to sleep, everything stops.
The second is renting a small cloud server. The author's reasoning is straightforward: if you want that four a.m. message, the program has to be awake while you sleep. He mentions that this kind of server starts at around 6 US dollars a month. That figure is spoken in the video, translated from the subtitle track, and not independently verified.
He also says you do not have to choose one or the other, since you can later connect the desktop app to that server and share the same agent.
The server in the video is from Hostinger, and the author uses its one-click template. He also says the first link in the description gives an extra 10 percent discount, which makes this a recommendation-type segment. This article does not include any link or discount code and does not compare hosting providers. You can use an environment you are already comfortable with. The author says the official documentation has instructions for running it for free on your own computer, which we have not read.
Next is the installation itself. The author sums it up in three steps: give it a place, a brain, and a phone.
He compares the whole process to hiring your first AI employee: a workplace, a brain, and a phone number. What follows only relays what the subtitle track states; for commands and interface details, follow the official documentation.
The place: after choosing a server plan, deploy with the template. The deployment form provides a default admin account (the author says to keep it as Hermes) and a password. The author reminds viewers to copy the password into a password manager, because every screen afterwards needs it. When deployment finishes you get a web control panel. The author warns that sometimes the setup screen does not open automatically after deployment, and you have to add it manually from the container management page in the server dashboard.
The brain: a freshly installed agent has no model, and the top right of the chat window shows that no provider has been set. The author uses an aggregator account such as OpenRouter to get access to multiple models, pastes in the API key, and sets the main model to DeepSeek-V4-Flash, because it is cheap, fast, and good at using tools. He also mentions that the official Nous Portal is a flat subscription starting at around 20 US dollars a month, but he starts with pay-as-you-go billing.
The phone: on the channels page of the control panel, choose Telegram and scan the QR code with your phone to create a dedicated bot. Save, restart, and you can start chatting. The author specifically notes that the QR code expires after about 3 minutes, so have your phone ready first.
The author estimates the whole process at about 30 minutes, with no coding required, just copy and paste. That time is spoken in the video, and in practice it will depend on your account verification and payment steps.
Monthly Cost and Brakes: The Bill Is Not the Server, It Is Re-Reading Everything Every Time
Most people look first at the monthly server fee. The author's demo points somewhere else.
On the OpenRouter activity page, he saw that during the recording of this video there were 105 requests costing about 13 cents. But the token count on the same page was striking: about 4 million, which works out to roughly 38,000 tokens for every single message.
The reason is that on each turn the agent sends more than your question. It sends its instructions, its memory, its tool list, and the whole conversation so far, all over again. This repeated content is called context, and the author says it makes up almost the entire bill.
So why only 13 cents? The author says about three quarters of the reads hit the cache, and providers charge very little for the parts that have not changed. At that rate, he estimates the initial 5 US dollar top-up could support about 4000 messages.
These are numbers from a single demo, translated from the subtitle track, and nobody can promise your usage will look the same. The author only states that a more expensive model makes the bill higher. Whether frequent web browsing or long-running tasks also increase usage is a reminder from this article, not a figure from the video. The server fee is fixed and the model fee floats, and that is the part to watch.
Set the brakes first.
The author suggests three practices, none of them hard.
First, top up a small amount. He suggests putting in 5 US dollars to test and not choosing the largest option at the start. Second, set a spending limit on that key. He demonstrates 10 US dollars per week, and names the key after its purpose so you can later tell who is spending. Third, make a habit of checking the activity page: how much each request cost and which model was used are all visible there.
On switching models, the author says to use a cheap model for daily work and switch to a stronger one for difficult tasks, and that the list includes free models with limited quotas to try. He also reminds viewers that the big vendors' flagship models cost more than the one he picked. In Telegram, he says you can switch with the /model command, which is the subtitle track's rendering of the command, so please confirm it in your own environment.
Security: It Is Not Answering Questions, It Is Taking Action
The author spends a good deal of time on security, and this is the part most worth pausing on as a beginner.
The agent writes files, runs programs, and browses the web on its own. In the author's demo, the agent installed a dependency package by itself in order to finish a scheduled task. Web content can also carry malicious instructions, known as prompt injection, aimed at getting the agent to do things its owner never intended.
The author's approach is to put it on a dedicated server with nothing else on it. In his words, that way it can reach at most that rented machine, not the personal documents on your computer. This is his design choice. If you install it directly on a laptop, the same abilities are pointed at your real files. He says he would rather let it learn on its own server.
A few other measures are worth noting, all as the author describes them. The setup flow adds your account to an allowlist, so if someone else finds the bot and messages it, they get no reply, and you can add people later. For actions such as downloading packages or deleting sensitive files, the agent stops and asks for your permission first. Treat the key like a password and do not share it.
To be clear: a dedicated server only narrows the scope. It is not a bulletproof vest. The protections the author describes have not been tested by us, and if you plan to let it touch company data or any account permissions, check with your security lead first.
Once You Are Comfortable: Schedules, Skills, and Other Features
The focus of the author's demo is scheduling: state the need once, ask it to build its own checking routine, keep a record of what it has already reported, run every few hours, and message you only when something new appears. The subtitle track shows he set it to every four hours, and the schedule page in the control panel shows the run count and next run time, and lets you pause or run manually.
The author also walks through a few more features, all of them the author's claims: it can dispatch sub-agents to handle split-up tasks at the same time; you can create multiple profiles to separate work from personal use; besides Telegram, channels can include Discord, Slack, WhatsApp, and email; and if you have used OpenClaw, you can ask it to help import your settings.
These touch on the integration of skills and tools. If you want to understand the standards behind them, see the further reading at the end.
FAQ
Q1: Does Hermes Agent itself cost money?
The video author says the software is free and that what you pay for is model usage and the server fee. That is the author's claim, translated from the subtitle track. Please check the official documentation for license and pricing.
Q2: Do I have to rent a cloud server?
Not necessarily. The author says the desktop app is the fastest way to try it, with the downside that it stops when your computer sleeps. You only need an always-on environment if you want automatic overnight schedules. The official documentation also has instructions for running it for free on your own computer.
Q3: Roughly how much will it cost each month?
The server in the author's demo starts at around 6 US dollars a month, and the model fee in one demo was about 13 cents over 105 requests. These are spoken figures and not independently verified. Actual cost will vary with the model, the tasks, and how much you use it, so set a spending limit first and then observe.
Q4: Is it safe to leave running?
The author designed in a dedicated server, an allowlist, and ask-before-dangerous-actions, but these only reduce risk. They are not guarantees. Assess for yourself before handling sensitive data.
Source Statement
This article is based on the Arabic subtitle track, not a transcript in the original language. Proper nouns, numbers, commands, and prices may contain translation errors, none have been checked against the original video, and none have been independently verified. All prices, usage figures, and capability claims in the video are attributed to the author and do not represent testing or endorsement by this site.
The video recommends Hostinger and offers an extra discount, which is a recommendation-type segment. This article includes no related links or discount codes.
Reference source: the video titled Hermes Agent - Full Tutorial & Setup Guide (For Beginners), channel Metics Media, published on August 31, 2026, original link https://www.youtube.com/watch?v=DYdvJCxWd6M. This article is a reorganization with added explanation, not a sentence-by-sentence translation.
Further Reading
If this topic interests you, we have published related articles on this site: what Meta Muse is and how a personal agent compares with OpenClaw, what Agent Skills are as an open standard in 2026, what AGENTS.md is as an open standard in 2026, and why MCP went stateless along with the A2A agent architecture.
Try a Small Budget for a Week Before Deciding to Leave It Running
The bill for this kind of agent only shows itself after a few days of real use. Instead of loading it with tasks right away, pick a single scheduled job, for example checking a topic every few hours, run it on a small budget for a week, and then see which model cost how much.
If you want to know how much difference a different model makes on the same scheduled task, you can use a single key on AI Token King to run different models for a day each, write down the spending for each day yourself, and compare. If you would like to look first and decide later, you can learn more at https://aitokenking.com.tw/home
