Hermes Agent is an open-source AI agent from Nous Research. It first shipped on 25 February 2026 and has passed 240,000 GitHub stars since. It is not a hosted product. You run it on your own laptop or a cheap server, point it at whatever model you want, and talk to it through a chat app or the terminal. The one thing that sets it apart from every other agent: after a hard task it writes itself a reusable skill, so it keeps getting better the more you use it.
01. What Hermes Agent is
open source, self-hosted, model-agnostic
It is MIT licensed and built by Nous Research, with Teknium, one of the cofounders, as the lead. Nous runs 12 instances of it every day to build Hermes itself, which is a decent signal about how much they trust it.
You install it with one command. It connects to Nous Portal, OpenRouter, OpenAI, a local model, or your own endpoint, and you can switch between them. You reach it through Telegram, Discord, Slack, WhatsApp, Signal, or the CLI, all from a single gateway process. It reads and writes files, runs commands, searches the web, and schedules its own recurring jobs.
The pitch against the hosted agents is simple. No per-seat fee, no SaaS tier, no vendor holding your context. The full stack runs for about nine euros a month on a small VPS.
02. The five pillars
memory, skills, soul, crons, self-improvement
Everything Hermes does sits on five parts, and knowing what goes where is most of the learning curve.
| Pillar | What it holds | Where it lives |
|---|---|---|
| Memory | Facts about you and your setup. Project paths, preferences, environment. Bounded on purpose, around 2,200 characters. | ~/.hermes/MEMORY.md |
| Skills | Procedures. Step-by-step how-tos the agent loads only when a task needs them. | ~/.hermes/skills/ |
| SOUL.md | The agent's personality and default voice, so it sounds the same across sessions. | ~/.hermes/SOUL.md |
| Crons | Scheduled jobs written in plain language. Each one runs in a fresh session with no memory of your chats. | set with a sentence |
| Self-improvement | A background review after a session that proposes new or edited skills, with an approval gate you can turn on. | the skill loop |
The rule the docs repeat: memory is for what, skills are for how. A fact goes in memory. A workflow goes in a skill.
03. The skill loop, the part that matters
it writes its own procedures, and some are wrong
Roughly every 15 tool calls on a real task, Hermes pauses and writes a skill: a markdown file that turns what it just did into a recipe. Next time a similar task comes up, it loads that file instead of working the whole thing out again. This is why a Hermes instance on day 30 can be sharper than a fresh one, and it is the reason to pick this agent over a stateless one.
It also writes bad skills. Its own review loop, the Curator, tends to think it did well even when it did not. So the loop only pays off if you keep the folder clean. Open ~/.hermes/skills/ once a week, read the new ones, delete the two that got it wrong, fix the one that is nearly right. Five minutes. Run any new skill three to five times before you depend on it.
The agent on day 30 beats day 1 only if the skills it kept were the right lessons.
The format follows the agentskills.io open standard, the same one other agents use, so skills are portable. There is also a separate companion project, GEPA, that reads failure traces offline and evolves better prompts. That one is advanced and optional. The setup and skill-loop guides below go deeper.
04. What it runs on, what it talks to
seven execution backends, six ways to reach it
Hermes can run commands in seven places: your local shell, Docker, SSH, Daytona, Singularity, Modal, and Vercel Sandbox. Use Docker for anything untrusted by setting TERMINAL_ENV=docker, and code it runs cannot touch your host.
On the input side it listens on Telegram, Discord, Slack, WhatsApp, Signal, and the CLI at the same time. Set a home channel with /sethome and that is where scheduled job results land. There is a GUI profile builder in the desktop dashboard for anyone who does not want to edit config by hand.
One thing to know early: switching models mid-conversation breaks the prompt cache and forces a full re-read at input price. Pick a model per session rather than hopping.
05. What it costs
free agent, you pay for tokens and a box
The agent is free and MIT licensed forever. Your only costs are model tokens and the machine it runs on. Nous Portal exists so you can start without collecting a wall of API keys.
People running it in production mostly use 70B-class or smaller models for cost and speed, and reach for a frontier model only on the hard reasoning. One documented lead-generation job ran on a server costing 24 cents an hour. Another user runs the whole stack, VPS and deployment and containers, for around nine euros a month.
06. Install it
one command, then the setup wizard
# macOS / Linux / WSL2
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
# Windows (PowerShell)
iex (irm https://hermes-agent.nousresearch.com/install.ps1)Then run the wizard, reload your shell, and start the CLI.
hermes setup --portal # logs in, sets the provider, turns on the tool gateway
source ~/.zshrc # or ~/.bashrc
hermes # start talking to itIt is a curl-to-shell install. If that bothers you, open the script in a browser first and read it, same as you would with any install line. The setup guide in this set has the full first-run walkthrough, including model choice, wiring your first chat app, and the parts that trip people.
07. Watch it explained
one talk on the why, one walkthrough on the how
Nous Research presented the case for running open models in production at the Arize Observe 2026 conference. It is the clearest short explanation of why Hermes is built the way it is.
Arize Observe 2026. External video, opens on YouTube.
For the hands-on version, this first-look walkthrough installs Hermes from scratch and runs the quickstart, so you can see the setup flow before you touch a terminal.
A community quickstart walkthrough. External video, opens on YouTube.
08. Where it actually pays off
the grounded cases, from the official user stories
The honest wins are boring and repeatable. A freelancer in France charges 2,700 euros a month to install and adapt Hermes for small businesses, and says the value is in the adapting, not the install. One user is on day 297 of a daily automation streak and reports more than 100,000 dollars of client work value now automated. Gumroad runs support tickets, mention monitoring, and finance-doc updates through it on a dedicated Mac. Miles Deutscher connected his whole business, notes, SOPs and meeting logs, into it alongside an Obsidian vault.
Not all of it holds up. There are posts claiming fast trading returns from Hermes bots. Treat those the way you treat every trading claim online. The repeatable value is in summaries, monitoring, first drafts, and the grunt work of client delivery. The where-it-pays-off guide breaks the real cases down with the numbers.
09. The four guides in this set
start with setup, come back for the rest
| Guide | What it covers | Read |
|---|---|---|
| Setup and first run | The two install commands, hermes setup, choosing a model, wiring your first chat app, the first real task, and where it trips. |
read |
| The skill loop | How it writes its own skills, when to trust one, the approve gate, the agentskills.io standard, and GEPA. | read |
| Integrations and automations | The seven execution backends, cron jobs, daily briefings, posting to X without the paid API, and the community skill directory. | read |
| Where it pays off | The real documented cases, the numbers behind them, and an honest read on which ones are actually repeatable. | read |
10. The catch, and where to start
the loop cuts both ways
The skill loop is the whole point and also the main risk. Left alone, the folder fills with mediocre skills the agent trusts anyway, and month two is worse than week one. The five-minute weekly cleanup is not optional.
It also runs real commands on a real machine. Keep the dangerous-command check on. Never set GATEWAY_ALLOW_ALL_USERS=true on a bot that has terminal access. Use Docker for anything you did not write. And remember cron jobs run blind, with no memory of your chats, so every scheduled instruction has to stand on its own.
To start: install it locally, run hermes setup --portal, and give it one task you do every week. Let it write the first skill, then open that skill and read it. If it caught the right lesson, you will feel it. From there, wire one chat app and one cron job. The setup guide walks the whole thing.
Sources and further reading
- github.com/NousResearch/hermes-agent and the official docs. MIT, built by Nous Research.
- User stories, where the freelancer, day-297, Gumroad and Miles Deutscher cases come from.
- hermes-agent-self-evolution, the GEPA companion repo. agentskills.io, the skill format standard.
- awesome-hermes-agent, the independent community directory, and the Nous Discord archive.
- Teknium on X, Nous cofounder and Hermes lead. Nous Discord.
- Videos: the Arize Observe 2026 talk and a first-look walkthrough. Text tutorial: DataCamp.
// Free newsletter
I send out guides like this every week
Real setups, real sources, no hype. Drop your email and I'll send you the next one.