The Server I Forgot I Owned Now Runs My AI Team

Plain-English guide · 8 minutes · for when you’ve become the bottleneck.

Hermes Agent is Nous Research’s open-source, self-hosted framework that runs an AI agent on your own machine, lets it delegate to other agents and models, and gives it a memory that survives between sessions.

Who this is for: you’ve built a few AI skills, you have tasks sitting fully specified in a list, and the only reason they’re not done is that nobody has started them. Who should skip it: if you haven’t got one Claude Code skill working on your laptop yet, start with Claude Code for Non-Coders and come back.

After reading this you’ll know whether you need a team lead yet, what it actually takes to run one, and the checklist that keeps it from eating your server.

Figures last verified 2026-09-22 against my own install notes.

So what is it?

Everyone using AI right now is doing it one chat at a time. You type, it answers, you wait.

Hermes Agent turns that into a team. One agent takes a job, splits it, hands pieces to helpers, remembers what it learned, and keeps working when you close the laptop.

In their words, from the README: “The self-improving AI agent built by Nous Research… it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations.” And the line that sold me: “Spawn isolated subagents for parallel workstreams.”

Translated: it can hand work to helpers and run them at the same time. That’s the difference between a chatbot and a team.

Is it the same as the Hermes models?

No. Same company, different product. Nous Research also makes AI models called Hermes. Hermes Agent is the manager that uses models, theirs or anyone’s.

I assumed it was a model and spent my first hour researching the wrong thing. If you’ve read What Is DeepSeek Harness?, here’s the relationship: a harness like dsh or Claude Code is one worker at one desk. Hermes is the person who assigns the desks.

Do I need one?

Probably not yet. You need it the day you notice you’re the bottleneck for your own automations. Three signs:

1. Tasks sit in a list, fully specified, and nobody has started them.
2. You keep re-explaining the same context to every new chat.
3. You’d pay a part-time assistant if the work were more predictable.

Hermes is the assistant for the predictable part. The trick is deciding what’s predictable before the agent spends money finding out.

How do I run it without a big bill?

Four choices that kept mine near zero.

A server I forgot I owned. Last November I bought a tiny VPS in a Black Friday sale and never touched it. This September I found the welcome email. Two CPUs, a couple of gigabytes of memory. If you have an old droplet, a Raspberry Pi, or a mini PC in a drawer, you already own the hardware.

The tools I already pay for. Hermes needs a model. The default advice is to buy an API key. I did something cheaper: Hermes has a built-in skill that hands coding work to OpenAI’s Codex CLI, and Codex can sign in with the ChatGPT plan I already have. The builder on my server costs me nothing extra.

That choice has a catch, and most write-ups skip it. Letting an unattended server borrow a personal login sits in a gray area of most providers’ terms. I wrote down why I accepted the risk on the day I made it. The rule: if you can’t write down why, buy the API key. On my own machines, agents never borrow my Claude login. Ever.

A memory it shares with Claude Code. Hermes remembers between sessions. So does Claude Code, through a folder of markdown files with a small header on each. I made a plain git repository of those files and pointed both at it. Now the agent on the server and the agent on my laptop read the same notes. It’s a folder of text files. That’s the whole “shared memory” architecture.

A gate in front of it. Before Hermes spends a run on a task from my board, a tiny model answers one question: does this task have enough detail for an autonomous agent to act on unattended? No → back to me. That gate is Jev. Cheapest insurance in the system.

A container needs a real, long-running main process. Mine restarted 3,717 times because I gave Docker an interactive program to watch.

Where does it go wrong?

Three real bugs from my install. Each is a lesson for any server tool.

1. The restart loop. The container’s default command was the interactive text interface. With no terminal attached, that program exits instantly. Docker saw the main process exit and restarted it. 3,717 times. Each restart re-downloaded a small file into a temp folder that never got cleaned. The disk filled to 100% and everything on the box fell over. Check: the container’s main process is something long-running, never an interactive program.

2. The wrong drawer. I told Docker to save Hermes’s data at one path. Hermes actually stores everything at a different path. So my “persistent” storage had been persisting nothing since day one, and every rebuild silently wiped my logins. Check: after setting up storage, restart once on purpose and confirm the data survived.

3. The step I won’t delegate. Codex runs its work inside a sandbox that needs a permission from the host operating system. That’s a security change. I don’t let an AI make security changes to a server, so it’s waiting for me to type one command myself. As I write this, it’s still pending. Check: name the human-only steps before the agent asks.

Both fixes fit in the one command that starts the container. This is the shape that finally held:

docker run -d --name hermes-agent \
  --restart unless-stopped \
  -v hermes-data:/opt/data \        # Hermes keeps everything here, not ~/.hermes
  -p 127.0.0.1:9119:9119 \          # dashboard, reachable only through an SSH tunnel
  nousresearch/hermes-agent:latest \
  sleep infinity                     # a real main process; run `hermes` via docker exec

Three lines carry the lessons: the volume at the path the program actually uses, the dashboard bound to localhost so it’s never on the open internet, and sleep infinity so Docker has something long-running to watch.

One that isn’t a bug: Hermes writes its own skills over time. Powerful, and also something you have to read. Budget an hour a week to look at what it taught itself.

Do this in 10 minutes

1. 01
   Decide if you're the bottleneck. Count the tasks in your list that are fully specified and still not started. Under three: you don't need this yet. Bookmark it.
2. 02
   Find the computer you already own. An old VPS, a Pi, a mini PC. Two CPUs and 2 GB is enough. No GPU.
3. 03
   Copy the checklist below into your vault. Read it before you install anything. It's the whole week I lost, in seven lines.

Before an agent runs unattended
- [ ] Main process is long-running (not an interactive TUI).
- [ ] Storage path verified by restarting once and checking data survived.
- [ ] Model access: API key, OR a written note of why a subscription login is OK here.
- [ ] Never my Claude login. (Hard rule.)
- [ ] A gate in front: "enough detail to act unattended?" → else back to me.
- [ ] One human-only step named in advance (security changes, payments, sends).
- [ ] Shared memory = a git repo of markdown, same header format as my laptop agent.

Send this to the friend who just bought a server “for AI stuff.” The checklist is the part they need before the install.

Questions people ask

Is Hermes Agent free? Yes, open source under the MIT license. You pay for the model it uses and the server it runs on.

Hermes Agent vs Hermes model: what’s the difference? The models are AI brains. The Agent is the manager that can use any brain. Same company, different product.

Do I need a GPU? No. It’s a manager, not a model. Mine runs on a two-CPU box with no graphics card.

Can it use Claude? Yes, through an Anthropic API key. Never through your Claude subscription login.

Hermes Agent vs Claude Code vs DeepSeek Harness? Those two are single workers with tools. Hermes coordinates workers, including those two.

Is it safe? As safe as the boundaries you set. Use the checklist, and read their docs.

Where do I start if I’m not technical? Not here. Get one skill working in Claude Code for Non-Coders and come back when you’re the bottleneck.

Sources and further reading

• Hermes Agent — GitHub repo · documentation. The “self-improving” and “subagents” lines are theirs.
• Nous Research — nousresearch.com, the lab behind both the Agent and the Hermes models.
• Docker’s restart policy docs — why an exiting main process loops forever under --restart.
• The dispatch gate — how Jev sits in front of the agent, in the Jev technical post.

The tools I run, in plain English

• What Is Jev? — the $0.0002 yes-or-no AI that gates everything.
• What Is DeepSeek Harness? — the open-source Claude Code alternative and the off-peak trick.
• Jev in front of Claude Code — the technical build log, including the dispatch gate.
• AI for Small Business — where all of this fits in a real week.