agentsasfolders · blog

Hermes Agent, Explained By Someone Running A Fleet

Jordan Shaw · August 28, 2026

Search interest in Hermes Agent went from almost nothing in January to hundreds of thousands of searches a month by spring. If you just typed it into Google and landed here, this page is the answer I wish existed when I started: what it is, who makes it, and what it's like to actually run. I run a fleet of private Hermes agents in production, so this is written from the operator's chair, not from a press release.

Who makes Hermes Agent?

Nous Research. They started as an open source AI lab known for the Hermes model series, and TechCrunch reported in July 2026 that they were raising new funding at a $1.5 billion valuation. The agent runtime, the Hermes Desktop app, and the Nous Portal subscription are all theirs.

The license is MIT. Said another way, the code is genuinely yours to run, modify, and keep, and nobody can take the runtime away from you or start charging rent on it later. In a year where AI tools kept getting repriced and shut down under their users, that detail stopped being a nerd preference and became the reason people pick it.

What does Hermes Agent actually do?

It runs missions, not messages. You give it a job, it works the job across steps and tools, and it carries memory between sessions. It can also write its own skills as it learns your work.

That memory is the part operators talk about. One Hacker News user put it better than any spec sheet:

"having a competent agent with constant state has been good for memorializing and organizing... I've been pretty pleased. Surprised even, if I'm being honest."

Constant state, said another way: it remembers who you are and what you were doing yesterday. ChatGPT's memory tops out around 1,300 words and fills up silently. A Hermes agent's workspace is files on disk. It doesn't forget you overnight.

The 3 ways people run it

1. The Desktop app. Hermes Desktop shipped in public preview in June 2026. Free, no account required, same agent core as the command line version.
2. Self-hosted. You put the runtime on your own machine or server. Full control, and full responsibility for setup, updates, and security.
3. Hosted for you. Somebody stands up an isolated instance on their hardware, locks the front door, and hands you the keys. This is what I do, 1 agent per person, each in its own sealed room behind Cloudflare Zero Trust. Said another way, your agent gets its own address and its own login, and a wrong email at the door gets silence, not an error.

The honest part: it's a v1

Hermes Agent is young and it acts young. The complaints you'll find on Reddit are real. It sometimes grades its own work too kindly. Its release cycle is immature next to older projects. It has fewer integrations than OpenClaw, which is the other name you'll keep seeing. I wrote a straight Hermes vs OpenClaw comparison if that's your actual question.

None of that changes the direction. The most-used agent runtime in the world, open source, backed by a well-funded lab, with memory that compounds. The rough edges are the price of being early, and being early is the entire opportunity.

What should you do with one?

The mistake is asking it to run your business unsupervised on day 1. The move that works is the night shift: give it a mission in the evening, let it draft and prep overnight, and have it stop at approval gates so nothing gets sent, spent, or deleted without your yes. You review a packet over coffee.

That pattern is written up as a free playbook, and it works on the AI subscription you already pay for. If you want to see other people's setups and ask questions, the free community is at Skool.