Logo ezacae
All articles

// blog · Hermes Agent · AI agent · agentic AI · n8n

What is Hermes Agent and what is it used for in business?

By , developer at ezacae · October 26, 2026

I installed and configured Hermes Agent at ezacae, where it runs our LinkedIn prospecting agent.

In short: Hermes Agent is an open source AI agent created by Nous Research. It runs on your server, remembers your conversations and writes its own skills so it does better the next time. The software is free; you pay for the AI model it calls and for hosting. At ezacae, it drafts our LinkedIn prospecting messages every morning, and nothing goes out without human approval.

You have seen the name everywhere since the spring and you are wondering whether it is for you. Here is what Hermes Agent is, what it really costs, when a classic workflow is the better choice, and how to stay in control. With the numbers from our own setup, break-in period included.

$0 for the software, open source under the MIT license, according to the official Nous Research website (accessed October 21, 2026)

€4 to €9/month excl. VAT for a small starter server, according to OVHcloud and Hetzner prices checked on October 21, 2026

What is Hermes Agent?

Hermes Agent is an autonomous, open source AI agent developed by Nous Research, an open AI research lab. Its promise fits in its tagline: “the agent that grows with you.” It keeps what it has learned from one session to the next and turns what it got right into reusable skills.

The code is published on GitHub under the MIT license: you can install it, modify it and use it in your business without paying for a license. The first release came out on March 12, 2026; version 0.22.2 was released on October 15, 2026. According to GitHub, the project had 255,696 stars on October 21, 2026, a sign of massive interest among developers.

The official website is hermes-agent.nousresearch.com. Other websites use the name with no link to Nous Research: download only from the official website or GitHub.

How does Hermes Agent work?

Hermes Agent combines four building blocks: an AI model that reasons, tools to take action, a lasting memory, and skills it writes itself. The fourth one is what sets it apart.

  • The model. Hermes has no AI of its own: it calls the model of your choice, from more than 40 providers listed in its documentation (Anthropic, OpenAI, Google, Mistral, DeepSeek…), or a model installed on your own machines.
  • The tools. More than 40 built-in tools according to the official README: web search, automated browsing, image reading, terminal, scheduled tasks, sub-agents working in parallel.
  • Messaging apps. You talk to it from Telegram, Slack, WhatsApp, Discord or Signal, or from the Hermes Desktop app (macOS and Windows).
  • Memory. It saves on its own what is worth remembering and can search your past conversations.
  • Skills. A skill is a how-to sheet the agent rereads when a task calls for it. According to the documentation, it creates one after a multi-step job worth repeating, after finding the fix for an error, or when you correct the way it works. It also updates its existing skills as it goes.

By default, it writes these skills without asking for permission. A setting lets you hold every change until you approve it.

“When you correct it, it updates its skill on its own. When you ask it for something it has never done, it writes a skill so it doesn’t start from scratch next time. That’s what I find most relevant.”

Patrice Dornon, developer at ezacae, October 2026 (translated from French)

According to their documentation, Hermes Agent is the only one that creates and corrects its own skills, especially when you correct it, compared with Claude Code and Codex, two widely used coding assistants. Those two tools also keep things from one session to the next, but as memory notes; creating a skill remains up to the user.

What is agentic AI, and how is an agent different from a chatbot?

Agentic AI is AI that does more than answer: it chains several steps on its own, with tools, toward a goal. A chatbot waits for your question and answers it. An agent receives a mission, decides on the steps, acts, checks and starts again.

Between the two, there is a third option many people forget: the automated workflow, a series of fixed steps you designed in advance.

Chatbot (ChatGPT, Claude…) Automated workflow (n8n…) Agent (Hermes Agent…)
Who decides the steps? You, with every message You, once and for all The agent, based on a goal
Acts on your tools? Little or not at all Yes, always the same way Yes, adapting its actions
Remembers from one run to the next? Depends on the tool No, it replays the same pattern Yes: memory and skills
Predictable result? Variable Yes Variable, needs guardrails
Best for Thinking and writing with you Repetitive tasks with clear rules Reading, judging and writing without you

Comparison by ezacae, October 2026, based on each tool’s documentation.

What can Hermes Agent do for a business? Our real use case

Hermes Agent is useful for tasks that require reading, judging and writing, over and over: monitoring, sorting requests, drafting replies, writing reports. At ezacae, it has been running our LinkedIn prospecting agent (in French) since October 2026.

Every morning, the setup spots posts where someone is looking for a service provider. A specialized decision model, Jev, does a first sort. Running on Hermes Agent, it rereads only the borderline posts and writes a short message in the tone of the post. Marie Lewin, our CEO, approves, edits or ignores it. Nothing goes out without her click.

Here is what it delivers, according to our measurements for the week of October 7 to 13, 2026:

  • 215 posts read on Marie’s behalf, 38 reread by the agent, according to our measurements;
  • on October 12, a message ready 12 minutes after the post was published;
  • about ten days to build it, and less than 10 minutes a day of Marie’s time (our estimate, during the break-in period).

Our first mistake is a useful lesson. With the question “Is this post relevant?”, the sort kept 9 posts out of 43, 7 of which were outside our offer. We replaced it with five closed questions (real request? what kind of work? within our offer?…). An agent does what you ask: if the question is vague, so is the answer.

It also has deliberate limits. It only cites what our offer sheet says: when Marie wanted to add skills that were not listed there, it refused and she rewrote the message herself. The hardest part was not the AI but the search terms: the best posts never say “looking for a service provider.”

Is Hermes Agent free? How much does it really cost?

The software is free, but a running agent costs three things: the AI model, the server and the setup time. For a well-chosen first task, expect a few euros to a few dozen euros a month, not counting people’s time.

Item Range Source
Hermes Agent software $0 (MIT license) According to the official Nous Research website, October 2026
Starter server (2 to 4 virtual CPUs, 4 to 8 GB of memory) €4 to €9 per month excl. VAT According to OVHcloud and Hetzner prices checked on October 21, 2026
AI model through the Nous Research plans $0 to $200 per month depending on the plan According to the official website, Free, Plus, Super and Ultra plans, October 2026
Setup About ten days for our first agent Our setup, October 2026

A few honest caveats. Nous Research does not publish minimum requirements: the 4 GB memory recommendation comes from hosting providers’ guides, not from the publisher. The lowest OVHcloud price requires a 12-month commitment, and the cheapest Hetzner plans were unavailable on the day we checked.

The hidden cost is the tuning: picking the right questions and search terms, and reviewing the first results.

Want an agent like this one for your team? Let’s talk it through in a 30-minute discovery call: the task to hand over, a realistic cost, the guardrails.

Hermes Agent or an n8n workflow: which one should you choose?

Choose n8n for anything deterministic, meaning when the same inputs must always produce the same result. Choose an agent when the job means reading a text, judging a situation or writing. And often, combine the two.

An n8n workflow is perfect for copying an order into your ERP, sending an invoice reminder 30 days after the due date or syncing two databases. The rules are clear, the result can be checked, the cost is fixed. Adding an agent would bring unpredictability for no benefit.

An agent becomes useful when the task cannot be described as “if this, then that.” Deciding whether a LinkedIn post is a real request, within our offer, then writing a reply in its author’s tone: no workflow can do that.

Our setup mixes both. Hermes reads and writes. But the pre-send checks, set according to our October 2026 rules (300 characters at most, nobody ever contacted before, fewer than 20 messages a day), are handled by a fixed program the agent cannot change. What must be guaranteed does not depend on an AI.

If you are unsure about n8n itself, we wrote up the 5 reasons not to choose it, and the cases where it is exactly what you need.

Is Hermes Agent safe for a business, and how do you stay in control?

An agent that acts on a server can also get things wrong there. The main risk is not that it “goes rogue,” but that it runs a bad command, reads a booby-trapped instruction hidden in some content, or sends something nobody has reviewed. Hermes Agent provides guardrails; it is up to you to configure them and add your own.

According to the security documentation, Hermes offers three modes for commands considered dangerous:

  1. Smart (default): a second model assesses the risk, refuses what is clearly dangerous and asks for your approval in borderline cases.
  2. Manual: it always asks for your approval before a dangerous command.
  3. Off: no checks at all. Keep it for an isolated test environment.

Whatever the mode, a list of catastrophic commands stays forbidden, with no possible exception. The agent can also work inside a container, a box isolated from the rest of the server. On messaging apps, nobody can talk to it without being authorized: by default, every stranger is refused.

Add three common-sense rules:

  • A human approves every action toward the outside world. At ezacae, no message goes out without Marie.
  • No access keys within the agent’s reach when it does not need them, and only an official installation source.
  • Regular monitoring. We added a control agent that checks every week for server and Hermes Agent updates, as well as known security vulnerabilities. Every Monday morning, it sends us a report on Google Chat with the tasks it recommends.

One last point: the content the agent reads often contains personal data, names in posts or emails. Data protection law applies (the GDPR in Europe): decide from the start what the agent keeps, and for how long.

Key takeaways

Hermes Agent is a free, open source AI agent that runs on your infrastructure and improves by writing its own skills. Its cost is mostly the AI model, a small server and the tuning time. It shines on tasks that require reading, judging and writing; for everything else, an n8n workflow remains simpler and safer. And it does not replace human approval.

Do you have a task that might suit it? See how we integrate AI agents, or talk to us directly in a 30-minute discovery call.

Sources

Article written with the help of AI, reviewed and approved by the author.

A project in mind?

Let's talk!