Most of the AI tools you use disappear the moment you stop talking to them. As soon as you close the tab, ChatGPT is gone. Quit the terminal or IDE, and Claude Code goes with it. They're useful while you're there, but they're not really running. They wait. For you.

Hermes Agent, from Nous Research, is the opposite of that. It's an open-source AI agent that is designed to stay online while you sleep, and quietly accumulates context the longer you use it. You message it from Telegram, Discord, Slack, wherever you already are. It remembers what it learns. It writes its own skills as it goes. And once it's set up, it works while you're doing other things.

If you are familiar with Karpathy's LLM Wiki post, this is the other half of the same shift. Karpathy's idea was that knowledge should be compiled and persistent rather than a rediscovery approach each time you open a chat. The principles applied by Hermes are similar for actions. The workflows are compiled into reusable skills, a context that compounds across sessions, and an agent that doesn't start from zero.

Before setting it up, consider how your workflow changes when you have an agent that's always on.

What Hermes actually is

Hermes is what people are calling a personal AI agent: something between an assistant, a server, and a small piece of automation infrastructure. The way Nous Research frames it on their site is that it's deliberately not a coding copilot bolted onto an IDE, and not just another wrapper around a single model's API. It's something that runs on its own machine, accumulates context as it goes, and becomes increasingly useful the more you use it.

There are a few things that make it different from what most people are used to:

It has a closed learning loop. When it completes a task, it can save the approach as a reusable "skill". This is a markdown file describing how it solved or can solve the problem repeatedly. Ask it to do something similar later, and it pulls that skill instead of figuring everything out from scratch. Over weeks, it builds up a library of patterns for the way you uniquely work. It is this part which is making people call it self-improving, although it is worth emphasising that this is a mechanism, not magic.

It has persistent memory across sessions. Most AI tools forget you between conversations. Hermes keeps a SQLite database of every session, searchable months later, alongside structured memory files about who you are and how you work. Ask it something three weeks later, and it can pull context from then.

It's model-agnostic. You can run it on Claude, GPT, Gemini, an OpenRouter key that gives you 200+ models, or local models via Ollama. Switching is one command. You are not locked into one provider.

It lives in your messaging apps. Telegram, Discord, Slack, WhatsApp, Signal. The agent stays the same; the inbox changes. You message it from your phone like you'd message a person. If you are frequently on the move, this is liberating.

It runs scheduled jobs. Every Monday at 9 am, summarise something trending for you.

The mental model

Think of Hermes as a small piece of infrastructure rather than an app. You're not opening a thing to use it; you're standing up a service. That's the part that throws people who are used to ChatGPT.

There are a variety of setup options, from running it on your machine to using a separate machine, such as a Mac mini. I opted to try Hermes on a VPS to ensure everything was isolated. To do this, you spin up a cheap virtual private server, €3–12 a month from somewhere like Hetzner, Hostinger, or Contabo. You install Hermes with a single command. You connect it to a model provider via API key. You set up a Telegram bot so you can message the agent from your phone. If you want to go further, there are a couple of optional extras. Tailscale connects your devices to a private network so you can reach your VPS from anywhere without exposing it to the open internet.

That's the whole architecture. Server, agent, model, messaging app, optional secure tunnel. Once it's running, you mostly just talk to it.

What it actually does for you

Some of the use cases I've seen people running:

A morning briefing that lands on Telegram before you wake up, the top three things in your industry, a summary of yesterday's emails, and what's on your calendar.

An email triage agent that reads your inbox overnight, unsubscribes from junk, archives newsletters, and shows you a digest of what actually needs a reply.

A research agent you delegate to.

A weekly skill audit. The agent reviews what you've asked it to do, identifies repetitive tasks, and suggests automating them as scheduled jobs. Over a few weeks, what was routine work may become automated.

How it compares to OpenClaw

Compared to OpenClaw, the agent platform Hermes is most often compared to, the trade-offs come down to philosophy. OpenClaw is broader, with more community-built integrations, and can be connected to a dozen channels. Hermes is more opinionated about memory, skills, and the learning loop. The community consensus is increasingly that they're complementary rather than competing; some people run Hermes as the planning layer on top of OpenClaw's tooling. But for most people who don't already have an agent running, Hermes is the easier place to start, since the defaults are sensible and the setup wizard handles most of the work.

The downsides

A few things worth knowing before you set this up.

It's still early software. There are multiple releases a month, so things occasionally break. You're expected to run hermes update regularly.

Costs are real and easy to lose track of. The agent itself is free, and the VPS runs €3–10 a month, but the model API calls are metered. Running Claude Opus on every task will get expensive fast. The fix most people end up with is OpenRouter — a single API key that gives you access to 200+ models.

Security needs your attention, and this cannot be stated enough. The agent has full access to a server. It can run commands, edit files, and browse the web. If you give it your real email and Google Drive credentials on day one, you're trusting a system that can make mistakes. The standard advice, and it's good advice, is to start with throwaway accounts, keep them isolated from anything sensitive, and add capabilities incrementally as you build trust. Run it inside Docker, lock down the firewall to allow only Tailscale and SSH, and never run it as root.

Hallucinations happen. You have to review what it builds, especially in the first few weeks.

Why it matters

AI is moving from something you use to something that runs. A tool you summon versus a service that's always on. The distance between those two ideas is bigger than it looks. And the question of what could be happening in the background while you are doing something else has become more real.

The shift Hermes represents is worth paying attention to even if you don't end up running it.