Hermes Agent

A chatbot with hands

Most AI assistants are chatbots: you type a question, and they reply with words. Hermes Agent is an agent — a chatbot that can also act. Give it a task in ordinary language and it can run commands on a computer, create and edit files, search the web, and work through multi-step jobs on its own, checking its own progress as it goes.

It's built by Nous Research and is open source (MIT-licensed). It belongs to the same family as tools like Claude Code and OpenAI's Codex — but where those are tied to one company's models, Hermes works with whichever AI "brain" you choose, and it runs on hardware you own.

How it works

Using Hermes feels like chatting, but behind the scenes it follows a simple loop:

1. You ask for something in plain language — "check this price every day and tell me if it drops".
2. It plans — it works out what steps the task needs, in what order.
3. It acts — using its tools: the terminal, files, web search, and more.
4. It checks — it verifies the result rather than assuming it worked.
5. It reports — and it can reach you wherever you are.

It can also run jobs on a schedule (like a price watch that fires every morning), and for big tasks it can spin up smaller "helper" agents to work on parts in parallel, then gather their results.

Memory — it actually remembers you

Most chatbots start every conversation from scratch. Hermes keeps persistent memory: who you are, your preferences, details of your setup, and lessons it has learned. Every new session starts with that knowledge already loaded — you don't have to repeat yourself.

Crucially, that memory is human-readable: plain notes and files you can open and edit, plus a searchable history of past conversations. It isn't a black box of inscrutable data. Memory is also pluggable — it can store what it knows in different backends, and separate "profiles" keep different contexts fully isolated from each other.

Skills — it learns on the job

The most distinctive part of Hermes is skills. When it figures out how to do something well — say, adding a page to this very website — it can save that procedure as a reusable "skill". Next time the same kind of task comes up, the skill loads automatically and the job is done faster and more consistently, with the mistakes of the first attempt already learned out.

That makes Hermes self-improving: the more it works with you, the better it gets at the things you actually ask it to do. Skills are stored as readable files too, so you can see exactly what it "knows" and edit it.

Where you can talk to it

One agent core, many surfaces — you use the same assistant everywhere, with the same memory and skills:

It runs on Linux, macOS and Windows, and because it's a gateway-style agent, the same instance can answer you on the desktop and ping you on a messaging app from a scheduled job.

Bring your own brain

Hermes doesn't force you into one AI provider. It works with any of the major ones — OpenAI, Anthropic, Google, DeepSeek, xAI, OpenRouter, and more — including local models you run on your own hardware. You can swap models mid-task, and if a provider goes down or runs slow, it can rotate through credentials automatically.

• Any LLM provider — no vendor lock-in
• Local models supported, for privacy-critical work
• Swap or upgrade the "brain" without changing the assistant

My setup

I run Hermes on my Raspberry Pi 5 home server, alongside the other services behind davn.uk. All of its memory and data live on my own hardware, and I talk to it mostly from the desktop app. This very page was written by it, working from a short brief — which is the honest demo of what an agent does.

Find out more