Hermes Agent: the open-source AI agent that lives on your server

Hermes Agent is an open-source autonomous AI agent built by Nous Research that runs on a server you control. Unlike a chatbot, it has persistent memory, can schedule its own work, run code in sandboxes, and control a browser - and because you self-host it, the data never leaves your perimeter. We use it internally at Nordspike for research, scheduling, and lightweight automations. Below: what it actually does, the six features that matter, and where it fits in a service business stack.

Agent vs chatbot: the actual difference

A chatbot is a window. You open it, type, it responds, the window closes. The state vanishes. Tomorrow you start over. That is fine for "summarize this PDF" but useless for "keep an eye on competitor pricing this quarter and surface anything weird."

An agent is a process. It lives somewhere. It remembers. It can schedule itself. It can call tools. It can fail, retry, and report. The difference is the same as between a temp who forgets your name on Monday and an employee who reads your last six months of decisions before answering a question.

Hosted assistants - ChatGPT, Claude, Gemini - are improving rapidly on this. They have memory features now. But the agent still runs on someone else's infrastructure. Your data is in their model context. For a lot of European service businesses with regulated client data, that is a non-starter.

The six features that actually matter

1. Persistent memory

Hermes keeps a long-running memory. Past conversations, decisions, and observations stay accessible across sessions. You can ask it about something it learned three months ago without re-explaining context. For service businesses, this is the difference between "an AI tool we tried" and "an assistant that knows our clients."

2. Scheduled automations (cron-style)

You can tell Hermes to do something on a schedule: every Monday at 9am, every hour, the first day of the quarter. This is what makes the "keep an eye on X" pattern actually work. The agent wakes up, checks, takes action if needed, logs what it did, sleeps. No human in the loop.

3. Subagents

For complex tasks, Hermes can spawn subagents - separate execution contexts for a sub-task. The parent stays in charge of orchestration; the subagents do focused work. This is how you avoid the "single 50-step prompt" failure mode that plagues most agent setups.

4. Sandboxed execution

When the agent needs to run code (data wrangling, file conversion, scripting), it does so in an isolated sandbox. Your host filesystem, secrets, and other processes are protected. This is the boring-but-critical feature that makes self-hosting safe enough to actually deploy in production.

5. Browser control

Hermes can drive a real browser - login flows, scraping, form-filling, multi-step navigation. Where API access exists, prefer the API. But for the long tail of legacy admin panels and partner portals where there is no API, browser control is the only way to automate.

6. Multi-platform integrations

Hermes integrates with chat platforms (Telegram, Slack, Discord), email, and the standard developer tools. You can talk to it from wherever your team already works. We have one running in a private Telegram channel that the operations team can DM at midnight when something needs to happen.

Where Hermes fits in a Nordspike stack

We do not deploy Hermes for client-facing chatbots - that is what hosted assistants are good at. We deploy it for the work that needs to happen between human touchpoints: market scans, competitive monitoring, internal research synthesis, scheduled reports, and the lightweight automations that would otherwise live as cron jobs maintained by the worst-case engineer who once heard about them in a meeting.

A typical pattern: Hermes runs on a small EU-hosted VM, has API keys for the client's tools (CRM, email, monitoring), runs scheduled tasks, and reports anomalies into a private Slack channel. It is not glamorous. It just works, week after week, and no data leaves the customer perimeter.

Reading more on the official site (with screenshots of the Telegram interface and the install flow) is worth your time: hermes-agent.nousresearch.com. The README and Telegram demo show the agent in action better than any blog post can.

Setup, briefly

Install is a single curl one-liner from the official repository, followed by hermes setup which walks you through the model provider, memory store, and credentials. Plan to spend an afternoon getting it actually production-ready: TLS, secrets in a vault, automated backups for the memory store, monitoring on the agent process. None of this is unique to Hermes, but skipping it is what turns "cool side project" into "outage waiting to happen."

When NOT to use Hermes

If your use case is "our salespeople want a smarter Notion," you do not need a self-hosted agent. Use a hosted assistant. Hermes pays off when at least two of these are true: the data is sensitive enough that hosted-by-default is uncomfortable, the workflow is recurring rather than one-off, the cost of a hosted SaaS subscription per user is starting to add up, or you want to compose multiple agents into something more ambitious than a single tool can do.

If you want this set up properly - on EU infrastructure, with the right monitoring, integrated into your client workflows - that is exactly the kind of work we do. See AI development partnership for how that engagement works.