Hermes Agent is an open-source AI agent that Nous Research released in February 2026 under the MIT license, so you can download it and run it yourself for free. The big difference from a chatbot like ChatGPT is that it keeps running: it takes your messages in chat apps such as LINE and Telegram, runs tasks on a schedule, and writes down how it did things as reusable skills. Correct it once and it adjusts, so the same kind of job gets easier each time. There are three ways to use it: run it yourself on a machine that stays on around the clock, use Nous Research's own Hermes Cloud, or let a managed service set it up for you.
Hermes models vs. Hermes Agent
Search for "Hermes AI" and you will see two different things. Nous Research has also published the Hermes family of open language models, but Hermes Agent is a separate project. It is not a model. It is a program that uses a model to get work done: it remembers you, keeps a schedule, drives a browser and connects to chat apps, while the actual reasoning comes from whichever model you plug in. OpenAI, Anthropic, Google Gemini and OpenRouter all work, so you do not have to use a Hermes model.
That also means there is no website where you simply log in and chat. Hermes Agent runs for long periods on a computer or server, and you talk to it through its web console, a terminal, or a chat app.
How it differs from a chatbot
A chatbot answers when you ask and forgets the session when you close the window. Hermes Agent behaves more like an assistant sitting at a desk. Tell it in LINE to "sum up yesterday's orders every morning at 8" and it creates a scheduled job, runs it on time and sends the result back to the chat. It can open a browser to look things up or fill in forms, and it has long-term memory across conversations: your preferences, your usual formats and the basics of your business only need to be said once. If you cannot remember what you discussed last month, it can run a full-text search over past conversations.
For chat apps, the project supports LINE, Telegram, Discord, Slack, WhatsApp, email and more. That matters in Taiwan, where most people already talk to family, coworkers and customers on LINE and would rather not install another app for an assistant. LINE does need a public HTTPS address to deliver messages, though, and if you host it yourself that part is on you.
How it gets better: skills, the Curator and memory
The learning loop is what people talk about most. According to the official docs, after a more complex task (roughly five or more tool calls) it writes the approach down as a reusable skill, and when it later finds a skill outdated, incomplete or wrong, it patches it. You do not maintain anything by hand; you just work with it normally and point out what to change. Ask it to turn orders into a weekly report, then add "include returns too", and the next report will include returns.
Skills pile up over time, so a background Curator keeps them tidy. It tracks how often each skill is viewed, used and patched, moves long-unused skills to stale and then archived, and periodically looks for near-duplicates to merge. Memory is a separate layer: short facts that are always useful, such as what to call you or the tone you like, sit in memory and are loaded at the start of every session, while longer procedures live in skills and load only when relevant.
By default the agent writes and edits skills on its own. If you would rather review every change first, the skills.write_approval setting holds new and edited skills for your approval before they take effect.
What people use it for
It shines on work that repeats every week, follows roughly the same steps, yet is never quite identical. A small shop or studio might have it compile a weekly order report every Monday, draft posts for new products, or send a reminder to reconcile accounts on the first of the month. A freelancer might have it sort incoming client emails and draft replies in a set format. After a few rounds it follows your corrected way of doing things, and you need to explain less and less.
On the other hand, if every service you need has an API and the flow is identical every time, for example "when a form is submitted, add a row to a spreadsheet and send a notification", an automation tool like n8n is cheaper and more predictable than an AI agent.
What self-hosting takes
There is an official Docker image, so self-hosting is not hard if you know your way around servers, but plan for a few things. First, it has to stay on 24/7; close the laptop or lose the home connection and the schedule and messages stop, so most people rent a VPS. Second, you set up the domain, reverse proxy and SSL yourself, the web console must have a login once it is reachable from outside, and LINE needs its webhook exposed publicly. Third, it moves fast: between mid and late October alone it went from 0.22.0 (October 14) to 0.22.5 (October 22), and every upgrade means pulling a new image. Fourth, model costs: you sign up for API keys with OpenAI, Anthropic or others and pay for what you use.
Then there is security. An AI agent touches your API keys, chat app tokens and whatever data you hand it, and the skills it learns are stored on the machine it runs on. Putting it on the computer you work on every day is like keeping the key next to the safe. A setup that is isolated from your work machine, with regular backups, is the safer choice.
Choosing between the three options
Self-hosting gives you the most freedom and suits people who are comfortable with servers and willing to handle updates and security. Nous Research's Hermes Cloud is another route: according to its official page, it runs an always-on Hermes Agent on a dedicated cloud instance with its own workspace and dashboard; you pick a size, deploy, and are billed daily from your Nous credit based on actual usage. The page lists Telegram, Discord, Slack, email and the CLI as ways to connect.
The third route is a managed service. RoamerHost, run by a team in Taiwan, hosts Hermes Agent in an isolated container for each member, with its own subdomain and SSL. The web console login is generated when the instance is created, the LINE webhook is already wired up, and it is ready about 60 seconds after payment. Pro is NT$599 per month and uses your own API key; the advanced plan is NT$1,599 per month and includes the Roamer AI model, so you do not need to sign up for any API key. Billing is monthly in Taiwan dollars with no lock-in, company invoices are available, and support is in Traditional Chinese. If you want to use it in LINE without running a server, this is the easiest path; plan details are on the Hermes Agent service page.
| Item | Self-hosted | Hermes Cloud (official) | RoamerHost managed |
|---|---|---|---|
| Where it runs | Your own VPS or computer | An official cloud instance | An isolated container at RoamerHost |
| How you pay | Machine costs plus model costs | Daily, from Nous credit, based on usage | From NT$599 per month, billed in TWD |
| Chat apps | Anything the project supports; LINE needs a public webhook you set up | Page lists Telegram, Discord, Slack, email | LINE (webhook ready), Telegram, Discord and more |
| Best for | People who know servers and want full control | People who use Telegram or Slack and are fine with an English interface | People who want LINE, Chinese-language support and TWD invoices |
Hermes Agent or OpenClaw?
The usual comparison is OpenClaw. Both are open-source AI agents, but they have different personalities: OpenClaw is a proactive personal secretary that watches your calendar, inbox and web pages and can run several agents with separate roles, while Hermes Agent is the assistant that learns you and gets better at repeated work. See OpenClaw vs. Hermes Agent for the full comparison.
FAQ
Is Hermes Agent free?
The software is MIT-licensed and free. What you actually pay for is the machine that keeps it running around the clock and the model usage, which depends on the provider and how much you use it.
Can I use it without coding?
Self-hosting means working with Docker, domains and SSL, which is a real hurdle if servers are new to you. With a managed service you just open the browser console or LINE and describe the work in plain language.
Where are its skills stored?
According to the docs, skills live in the ~/.hermes/skills/ folder on the machine it runs on, so they go wherever you host it; back them up before moving or reinstalling. On RoamerHost, memory, skills, conversation history and API keys are stored in your own container, isolated from other members.
WordPress
n8n
