Guides

The cheapest VPS that actually runs Hermes Agent 24/7

Published August 5, 2026 · Verified on v0.19.0

Hermes runs on your own machine. That is the whole point of it — persistent memory that lives on disk you control, no cloud lock-in, no telemetry. But it also means that if you want an agent that answers a Telegram message at 4am, something has to be awake at 4am, and a laptop that closes is not that thing.

This is the guide to picking that box without overpaying for it.

Disclosure: some provider links below are affiliate links, meaning this site earns a commission if you sign up through them at no extra cost to you. It does not change the recommendations — the reasoning for each is spelled out so you can disagree with it.

What Hermes actually needs

The single most common mistake is buying a GPU instance. Unless you are also serving the model yourself, you do not need one. With a hosted provider, Hermes is an orchestrator: it manages memory, runs tools, and shuttles requests. That workload is memory-bound, not compute-bound.

RAM is the constraint that bites. The Python runtime, Node.js for the tooling layer, a gateway process per chat platform, plus any MCP servers you add — each is modest alone and they add up fast. Memory grows with how many integrations you enable, not with how much you chat.

Disk grows quietly. Persistent memory is the feature you came for, and it accumulates. Twenty gigabytes is fine on day one and uncomfortable a year in. Pick a provider where resizing the volume is not a migration.

The shortlist

Hetzner — the price/performance answer, and it is not close. A 4 GB shared-vCPU instance lands in the €4–5/month range, which is roughly half what the US providers charge for the same specs. The catch is that datacentres are in Germany, Finland, and Ashburn, so pick accordingly. The control panel is spare compared to the competition. If you are cost-sensitive and comfortable on the command line, stop here.

DigitalOcean — you pay roughly double Hetzner for documentation that covers every problem you will have, a one-click firewall that is genuinely one click, and snapshots that work the way you expect. If this is your first server, the premium is tuition and it is worth paying.

Vultr — the answer when geography matters. Far more regions than either of the above, which is what you want if you are outside Europe and North America and care about SSH responsiveness. Pricing sits between the two.

Oracle Cloud Free Tier — genuinely free, genuinely enough (their Arm instances offer far more RAM than any paid tier at this price), and genuinely unreliable to obtain. Account approval is inconsistent by region and free instances can be reclaimed. Worth ten minutes to try; do not build a plan around it.

A Raspberry Pi you already own — zero marginal cost. A 4 GB Pi with an SSD runs Hermes fine when the model is remote. You trade a monthly bill for owning uptime, power, and your residential connection. For a personal agent that is often the right trade.

Skip these

Shared cPanel hosting cannot run Hermes — no persistent processes and no control over the Python environment. Generic function-as-a-service platforms are also the wrong shape for a long-lived gateway. Hermes does, however, support persistent cloud backends such as Daytona and Modal: they can hibernate compute while keeping the workspace available, so treat them as a separate option rather than ordinary “serverless functions.” Free-tier plans that simply kill state after inactivity still defeat the purpose.

Setting it up correctly

Create a non-root user first. Hermes installs into the invoking user’s home directory. Install as root and you get an agent living in /root that your normal user cannot reach — see the command not found fix for what that looks like when it goes wrong.

adduser hermes
usermod -aG sudo hermes
su - hermes

Verify before you integrate. Install, then:

hermes doctor
hermes chat -q "hello"

Both clean before you touch a single gateway. Debugging a Telegram integration on top of a broken model config wastes an evening.

When the base chat works, use the Hermes Agent Telegram setup guide for the BotFather token, user allowlist and first-message test.

Run it under systemd, not in a terminal. A gateway started in an SSH session dies with the session. First create the complete user unit from the Linux installation guide. Then, from an administrator account, allow the Hermes user service to survive logout:

sudo loginctl enable-linger hermes
su - hermes
systemctl --user enable --now hermes
systemctl --user status hermes

Firewall it. Hermes with chat gateways makes outbound connections; it does not need inbound ports open beyond SSH. Close everything else. An agent with filesystem and shell access is not something to leave reachable.

Snapshot before every upgrade. Hermes ships releases weekly and some of them break working setups. A snapshot turns a bad upgrade from an evening into ninety seconds. Check the changelog before pulling.

What this costs in practice

For a single-user setup on a hosted model provider: €4–10/month for the server, plus whatever your model usage costs — which will almost certainly exceed the server bill. Optimising the box below €5 while running an expensive model through it is effort spent in the wrong place.

FAQ

Do I need a GPU?

Not if you use a hosted model provider, which is how most people run Hermes. The agent orchestrates; the model runs elsewhere. You only need a GPU if you intend to serve the model locally too, and that is a completely different budget.

Is 2 GB of RAM enough?

For a bare CLI-only install you can sometimes get away with it. Once you add a gateway process and MCP servers, 2 GB becomes the reason things get OOM-killed at 3am. 4 GB is the practical floor.

Can I run this on a Raspberry Pi instead?

Yes, on a 4 GB or 8 GB Pi with an SSD rather than an SD card. It costs nothing per month and the latency is irrelevant when the model runs remotely. The tradeoff is that you are now responsible for uptime, power, and your home connection.