Hermes Agent setup: 7 things to set up right after install

I sent one message to my AI boss on Telegram. He handed the work to two other agents, they wrote at the same time on a Kanban board, and a few minutes later a Threads post and a short video script came back. All of it runs on a 4 GB server, with Hermes Agent and the Claude plan I already pay for.

This guide is the written version of my video on the 7 levels I set up right after installing Hermes Agent. You will build the same team: a boss you talk to on Telegram, two writers, and an ops engineer who checks your server every morning. Mine is a content team. You can build a research team, a coding team or a support team the same way.

Every SOUL.md file and every command from the video is in the companion repo: github.com/stackagen/hermes-agent-7-levels-setup.

Last checked: 2 November 2026, on Hermes Desktop connected to an Ubuntu 24.04 server.

⚡ Quick answers

What should I set up first after installing Hermes Agent? The desktop app, then a model. Hermes ships without a model, so nothing works until you plug one in. After that, profiles, a chat app, memory, skills and scheduled jobs.

Can Hermes Agent use my Claude Pro or Max plan? Yes, through the experimental Claude Subscription DirectSDK plugin. It uses your plan's limits, so busy agents eat into the same usage as your normal Claude chats. Do not pick the plain "Anthropic" provider, that one wants an API key.

What is a Hermes profile? A separate agent with its own personality (SOUL.md), memory, skills and settings. Think of it as a new hire with their own desk.

How do multiple Hermes agents work together? Through the Kanban board. A boss profile creates cards, worker profiles pick up their own cards, and the boss collects the results.

Do I need a server? No, Hermes runs on your laptop too. A server keeps the agents working when your laptop is closed, and that is what makes the morning report and the Telegram bot useful.

🧰 What you need

🗺️ The 7 levels at a glance

🖥️ Level 0: The desktop app

Hermes runs in the terminal, but there is also a desktop app for macOS, Windows and Linux. Download it from the Hermes Agent website.

My Hermes does not run on my laptop. It runs on a small server, so it keeps working when the laptop is closed. The desktop app connects to that server over SSH, which means the agent lives on the server and I control it from this window.

1. Open Hermes Desktop.
2. Click the backend switcher at the bottom left and add your server as a backend (host, user, SSH key).
3. Wait for the status bar to say Backend ready.

From here, everything in this guide happens in the app: profiles, plugins, files, scheduled jobs and even a terminal on the server.

My server is from AlphaVPS: 2 cores, 4 GB RAM and 40 GB storage, about 4 euros a month when I bought it. Check the current price before you buy. If you want the same one, this is my AlphaVPS link. It is an affiliate link, so I get a small commission at no extra cost to you.

🧠 Level 1: The brain

Hermes does not come with a model. You plug one in. Hermes is the body, the model is the brain.

I use my Claude plan because I already pay for it.

1. Go to Capabilities > Plugins and install and enable Claude Subscription DirectSDK.
2. Open the built-in terminal: press Ctrl+P (Cmd+P on a Mac), search terminal and toggle it on. This terminal runs on your backend, so on my setup it is the server.
3. Check that Claude Code is installed on the backend:

claude --version

If the command is not found, install it with the command from the Claude Code setup docs:

curl -fsSL https://claude.ai/install.sh | bash

4. Sign in:

claude auth login

Open the link it prints, authorize your account, copy the code and paste it back into the terminal. You will not see the code as you paste it, it is hidden on purpose. Press Enter and wait for the success message.

5. Go to Settings > Model, pick the Claude Subscription DirectSDK provider, pick a default model (I use Haiku 5.5) and apply.

Two things to know. The plugin is marked experimental. And it uses your plan's limits, so when the agents are busy, that is the same usage as your normal Claude chats.

You do not need Claude at all. In the same settings you can bring your own key for OpenAI, OpenRouter or the Anthropic API, or connect a local model. Just not on a 4 GB server like mine.

👥 Level 2: The team

Out of the box, Hermes is one assistant, the default profile. I keep that one as a normal chat.

Hermes also has profiles. A profile is a separate agent with its own personality, memory and skills. So I made an office. Yes, like the TV show.

For each one:

1. Create a new profile, name in lowercase.
2. Clone it from default, so it gets the same model, plugins and skills.
3. Add a one-line description.
4. Open its SOUL.md in the app, replace everything with the file from the companion repo, and save.

SOUL.md is the job description. It decides who the agent is, how it talks and what it does. Change the name "Aniket" to yours, and the voice rules to your own.

One line in Creed's file matters more than the rest:

- Never invent his stories, results or numbers. Write `[YOUR STORY: what]` or `[NUMBER]`.

A model that writes in your voice will happily make up a story about you. A gap you fill in is better than a story that never happened.

💬 Level 3: Telegram

Now the boss needs a phone number. Connecting a Hermes profile to Telegram used to mean creating a bot, copying a token and pasting it into a file. In the desktop app it is a QR code.

1. Go to Messaging, pick the michael profile and choose Telegram.
2. Pick Create with QR and scan the code with your phone. It opens BotFather and sends Start.
3. Create the bot and give it a name and a username.
4. Back in the app, click Connect this bot.
5. Check the allowed users list. It should hold your own Telegram user ID. To see your ID, message a bot like @userinfobot and compare the number.
6. Click Save and restart and wait for the gateway to come back.
7. Say hi to your new bot.

Why the allowed users list matters: Hermes can run commands on your server. If anyone who finds your bot could talk to it, anyone could run commands on your server. Keep the list to people you trust.

Do the same for dwight. He needs his own bot for the morning report in level 6.

📝 Level 4: Memory

Tell the boss what he should know about you, once:

Remember this about me: my audience is engineers and people who want to learn AI in really simple terms, with no hype. I post shorts every day. And I like hooks that start with a question.

He will say he saved it. A model saying it saved something and actually saving it are not the same thing, so check. Hermes keeps memory in small Markdown files per profile:

~/.hermes/profiles/michael/memories/USER.md
~/.hermes/profiles/michael/memories/MEMORY.md

In the desktop app, open the sidebar, open your home folder, then .hermes/profiles/michael/memories/USER.md. Your sentence should be there.

Memory is loaded when a new session starts. So in Telegram send /new, then ask Who is my audience?

What about Honcho or Mem0? Hermes supports external memory providers like Honcho and Mem0 as plugins, next to the built-in memory. I do not use one. For one person, the built-in memory is a few lines of rules and that is enough. Most providers send your messages to their service or need their own setup on your server, and on a 4 GB server I would rather keep the RAM for the agents. I would add one when many people talk to the same agent, or when I need to search months of old chats by meaning instead of keywords.

🧩 Level 5: A skill

Normally when a chat ends, the AI forgets how it did the job. Hermes can write down how it did a task as a skill and use it again next time. The agent writes its own recipe card.

1. Send the boss a draft with mistakes on purpose:

Review this draft before I post it: "So, I tested 12 AI agents last week — only one of them actually worked."

It has an em dash, a number I never gave, and a hook that is a statement instead of a question. In my test Michael caught all three.

2. Ask him to save it:

Save how you just reviewed this as a skill called review-drafts.

3. Check it in Capabilities > Skills, set Configuring to michael and search review. It is a plain Markdown file he wrote himself: what to check, in what order.

4. Test it with a new draft and no instructions:

Review this: "Basically, 9 out of 10 developers will switch to agents by next year."

He flagged the invented number and the statement hook without being told. Michael uses this same skill on Creed's and Kelly's drafts in level 7.

⏰ Level 6: A scheduled job

This one is Dwight's. His SOUL.md lists the read-only checks he runs: uptime, load, RAM, disk, the Hermes gateway, failed services, pending updates and whether a reboot is needed.

First, tell the bot where to deliver. Open Dwight's chat in Telegram and send:

/sethome

That marks this chat as the home channel. Scheduled job results go there. Without it, the report has nowhere to go.

Then create the cron job:

1. Switch the app to the dwight profile and go to Scheduled jobs > New.
2. Name: morning-health.
3. Prompt:

Morning check. Run your health checks and send me the report.

4. Deliver to: Telegram.
5. Frequency: Daily, at the hour you want. I picked 9 AM. Every 15 minutes is possible, but every run costs tokens.
6. Save, then trigger it once to test.

My first report said a reboot was pending. Dwight did not reboot anything. His SOUL.md says he can only read, and when something needs a fix he shows the exact command and waits for my yes.

🏆 Level 7: The boss level

One idea, the whole team.

Michael needs the Kanban board to hand out work and to know when it is done.

1. Turn on the board in the app: Ctrl+P, type kanban, enable the Kanban desktop plugin. It shows up in the sidebar.
2. Give Michael the Kanban tool on Telegram. I did not find this in the UI, so it is the one command in this setup. Run it in the built-in terminal:

hermes -p michael tools enable kanban --platform telegram && hermes gateway restart

3. In Michael's Telegram chat, send /sethome too, so his final pack has somewhere to land.
4. Put Telegram on the left and the board on the right, and send Michael an idea. Mine was a post about Google's new Genie world model.

What happens next, with nobody clicking anything:

• Michael replies in one line and creates two cards: one for Creed, one for Kelly. They start in Ready.
• Within about a minute the workers pick up their own cards and move them to Running. Both write at the same time on the same small server.
• Each card moves to Done and Michael gets notified.
• Michael reviews both drafts with the review-drafts skill. If a draft breaks a rule, he sends it back on a new card.
• The final pack lands on Telegram: the Threads post, the short script and which hook he likes best. He also appends it to ~/content/ideas.md on the server.

Each card keeps its description, summary, activity log and run log, so if one gets stuck you can see why and retry only that card.

Why not a group chat? I tried putting all the bots in one Telegram group first. The bots kept mentioning each other and Hermes stopped the group after a few rounds, which is what you want, otherwise they would go forever. For work like this, the board is the better tool.

⚠️ Gotchas I hit

• The plain Anthropic provider wants an API key. If you are on a subscription, pick Claude Subscription DirectSDK.
• Cron results go nowhere without a home channel. Send /sethome in the bot's chat first.
• Memory shows up after /new. A saved memory is loaded at the start of a session.
• Kanban in the app is not Kanban on Telegram. The desktop toggle is for the app. The terminal command is what lets Michael use the board from Telegram.
• Group chats with several bots stop after a few rounds. Use the board for multi-agent work.

🔒 The limits

The writers still get my voice wrong sometimes, so I read everything before I post. Nobody on this team posts anything, and that is on purpose. The Claude plugin is experimental. And whatever these agents do on your server is on you. Keep the allowed users list short and keep approvals on.

🧭 Where to go from here

• Every SOUL.md and command: github.com/stackagen/hermes-agent-7-levels-setup
• Prefer a sandbox on your own laptop? Running Hermes agents in Docker
• Official docs: Hermes Agent documentation

If something here did not work for you, tell me in the video comments and I will answer.