AI Agent On Your Own VPS 2026 — Hermes, Claude And Overnight Research

Short answer: to run an AI agent around the clock you need four things. A VPS for the agent to live on. The agent itself — mine is Hermes. A terminal whose sessions survive a disconnect, which is cmux. And a model that does the work. Everything else is convenience: Tailscale to reach the server from any device, Telegram to receive results, and voice dictation instead of typing. I walk through the whole stack on camera in „Zarabiaj jak śpisz z Claude" (Polish audio).

This post is the written version of that video. No scrubbing, timecodes to the parts that matter, and every unverified claim marked as unverified.

Disclosure: the video was made in cooperation with Hostinger and is sponsored. The Hostinger and Willow links here are affiliate links — if you buy through them I earn a commission, and the code DawidHermes10 gets you 10% off. That does not change what follows. The server I am describing has been running for months and this system genuinely lives on it. Check prices and specs on the vendor's own page, not here — they change faster than this post does.

Why this does not run on my laptop

This is the only reason worth reading further for. An agent on your own machine dies with the session. You close the terminal, close the lid, leave the house — the context is gone. Next time you explain everything from scratch.

On a VPS the session keeps running. In the video (04:28) I kill a session, open a new one, and everything is exactly where I left it. The context never ends. That sounds like a detail until you work with an agent daily, and then it turns out to be the whole difference between a toy and a tool.

Second thing: an agent on a VPS works while you do not. A laptop has to be on. A server always is. If you want something to happen at 3am, you need a machine that is awake at 3am.

In the video I use Hostinger. The plan on screen has 8 GB of RAM, servers can be placed in Europe, and there is a 30-day money-back guarantee. I am not copying the current price into this text because VPS promotions change monthly. The pricing page is binding, not this paragraph.

What the stack is made of

Component	What it does	Cost
VPS	the server the agent and its databases live on	paid — see vendor pricing
Hermes Agent	the agent itself: takes tasks, runs them, sends results back	install shown from 05:29
cmux	terminal with sessions that survive a disconnect	free
Tailscale	private network — reach your servers from any device	free on the plan I use
Claude	the subscription most of my work runs on	paid
Qwen	free model that took over research on my side	no usage fee
Telegram	the channel the agent reports to and I command from	free
Willow	voice dictation, works in Polish and English	see vendor pricing — I did not verify it

Three of these cost nothing, one is a server, one is a model subscription. There is no "agent platform" licence and no separate subscription for the agent itself.

Install: the agent sets itself up

This is the part people do not believe until they see it (05:29). You do not click through a wizard. You open the VPS panel, then hand your agent a link to the Hermes docs and tell it to install itself. It does the rest.

The same method keeps working. When I wanted a new model running on that server, I did not look for a tutorial. I pasted the model's link and described the outcome: install this, configure it on this server, then build me a system that runs research overnight. That is the entire "configuration".

You do not need to code. You need to describe precisely what should exist, and verify the result before trusting it. Same principle I wrote about in optimising Claude Code hooks: the agent executes, the human verifies.

I connect through Tailscale, so every machine I own shows up in one place with no public IP and no network configuration. Results come to Telegram because it is the simplest channel that exists. Hermes can hook into Discord instead if you prefer.

Most commands I dictate. The app is Willow, it handles both Polish and English, and I use it daily (06:45). Talking to an agent is simply faster than typing at it.

What changed: a free model took over research

The reason I filmed this the day after launch rather than a week later: a new free Qwen model had shipped the previous day (07:08).

I am deliberately not writing the version number into this text. Qwen models ship fast enough that the name from the video would go stale before this post does — the exact version is visible on screen in the video. On the benchmarks I show there, it comes out ahead of the older Opus generation. It is not better than Opus 5 and I am not claiming that. For the work I give it, it is more than enough.

Why this matters at all. AI subscriptions have usage limits, and the limit runs out exactly when you start working properly. My own example from the video: I am on the Claude 20x plan and at my usage 60% of the limit is gone with four days left before reset. An agent meant to grind 24 hours a day would burn through that in a day.

A free model does not have that problem. All research and store monitoring moved over to it, so I pay for the server rather than for each agent run. I did not drop Claude and I am not going to. Repetitive background work went to the cheap model, and the expensive subscription stayed for things that need a better head. It is the same split I use everywhere: expensive brain for decisions, cheap worker for the work.

What the system actually does while I sleep

The concrete part, and the body of the video from 07:33.

The agent has one job overnight: check what is happening in my market and have a report ready by 8am. It looks at competitors in my niches, at what changed in their offers, and at search trends. In the morning I get one document instead of opening fifteen tabs.

Example from the video (08:35). The report flagged a filter water bottle as something starting to climb. I open Google Trends on camera and check it live: the traffic genuinely only picked up in August. So the signal was real. Second item on the same list was a closet organiser.

And here is the line I will not cross. Google Trends shows interest in a search term. It does not show demand, it does not show margin, and it does not show profit. This is not investment or financial advice. A rising trend is a signal to investigate, not a decision to enter a product. Before you order anything you do normal product research and run the budget.

Before this, that report was either a person's job or mine, in the evening, after work. Now it happens without anyone. That is what the video title means — not that money appears on its own.

This is not a promise of results. Outcomes depend on the product, budget, execution, market and experience. Automating research shortens the time to a decision. It does not make the decision for you and does not guarantee it will be a good one.

What this system will not do

Four things stay on your side and no agent takes them over.

1. The product and offer decision. The agent brings a list of signals. Which one is a real opportunity and which is noise is your call, and your risk.
2. Creative. Understanding the customer and the angle is still the work of a human who knows the market.
3. Cashflow. A VPS and a subscription are fixed costs. If the budget does not work without them, it does not work with them either. Run the whole number first — I broke it down in how much dropshipping costs.
4. Verifying what the agent wrote. A model can be wrong with total confidence. For the first week, check every report by hand. If it disagrees with what you see yourself, find out why before you start trusting it.

If you are looking for something that runs a business from zero to profit on its own, this is not it and no such thing exists. Anyone selling that promise is selling a fairy tale.

Who this makes sense for, and who it does not

It makes sense if you do the same information work every day. Reviewing results, checking competitors, collecting trends, writing summaries. Those are hours per week you can hand to a machine. It also makes sense if you run several stores — work that is annoying with one becomes impossible with three.

It does not make sense if you do not have a store or first sales yet. At that stage money goes to the product and first ad tests, not infrastructure. Start with how to start dropshipping and stand up a server once there is something to automate.

It also does not make sense if you want a finished product with a dashboard and support. This is a box of parts you assemble yourself. If you would rather someone assembled it and wired it into your store, that is a different category of tool — I covered it in the AI tool for dropshipping.

How to test it in a week

Before you build anything large, build the smallest version and see whether you actually use it.

1. Day 1: stand up the VPS and install the agent. Nothing else.
2. Day 2: wire up Telegram notifications and send your first command from your phone.
3. Days 3-5: give the agent one repetitive task you do by hand today. One, not five. Check every result manually.
4. Day 7: decide. Can you name a specific job you no longer do? It stays. Cannot? The 30-day money-back guarantee exists so this decision is cheap.

That test works the same on my stack as on any other. I have no interest in you paying for a server that runs nothing.

If you want to see what the next step looks like, I post store builds daily on Instagram and run a weekly live call for the community.

FAQ

Do I need to code to run an AI agent on a VPS?

No. The agent performs the install itself — you give it a link to the docs and an instruction, and it stands up the rest. What you need is the ability to describe precisely what should exist and to verify the result. That is closer to working with a contractor than to programming.

How much does running this cost per month?

The VPS itself, plus optionally a model subscription. cmux, Tailscale on the plan I use, and Telegram are free, and I moved research to a free model so there is no usage fee for it. I am not quoting VPS figures in the text because promotions change monthly — Hostinger's pricing page is binding, not this post.

How is Hermes different from Claude?

They are two different layers. Claude is the model — it can reason and write. Hermes is the agent that lives on your server, takes tasks, runs them on schedule, and sends results to Telegram or Discord. Hermes can use different models, including free ones. An agent with no model cannot think, and a model with no agent cannot run anything overnight.

Is a free model enough to replace a paid subscription?

For repetitive background work, in my case yes, and that is exactly what I moved research to. For tasks that need stronger reasoning I still use the paid subscription. The cheapest split is: cheap model for the work, expensive one for decisions. Before you cancel anything, test the free model on your own tasks for a week.

Will the agent find a product to sell for me?

No. It will find signals: rising search terms, competitor moves, changes in offers. The decision about product, price and offer stays with you, along with the risk. A rising Google Trends line tells you about interest in a search term, not that money can be made on it.