Hermes Agent on a VPS: Best Plans and Setup in 2026

The short answer

Buy a 2 vCPU / 4 GB RAM x86 Linux VPS with 40 GB or more of SSD or NVMe. It is twice the recommended memory for a single-user Hermes Agent, so browser automation and busy turns never push the box into swap. You do not need a GPU. If you want the cheapest thing that still works, a 2 GB plan is fine for a chat-only bot. Everything below explains why, and what each provider charges today.

The same advice applies to OpenClaw, the other framework OpenClaw Launch hosts. Differences are covered in a dedicated section further down.

What Hermes Agent actually needs from a VPS

Hermes Agent, from Nous Research, runs on Linux, macOS, Windows and Termux. Its platform support page puts Linux and WSL2 on both x86_64 and aarch64 in the highest support tier, so an Arm VPS is a legitimate choice. On macOS, Apple Silicon is the top support tier, and Intel Macs are covered by a separate x64 desktop bundle and the CLI install — so an old Intel Mac can work as a home host, though a VPS is still the simpler way to stay online 24/7.

The sizing figures that matter come from the official Docker documentation:

The same page says to allocate at least 2 GB when browser automation is active, and that Playwright needs shared memory (--shm-size=1g). The image ships full Chromium, so the browser toolset costs disk as well as RAM. The installer also pulls in Node.js, ripgrep, FFmpeg and Chromium for you, so budget more disk than the 500 MB minimum suggests. We measured memory on our own fleet and cover it on the system requirements page: idle agents are tiny, and the figure to size against is the peak on a busy turn.

Our rule of thumb follows from that. Chat-only, one user: 2 GB. Browser toolset on, or any plan to run extra services on the same box: 4 GB. A 1 GB VPS technically meets the minimum, but it leaves nothing for the OS, Docker and a busy turn at once.

Why you almost certainly do not need a GPU

With an API provider such as OpenRouter, Anthropic or OpenAI, the model runs on the provider's hardware. Your VPS runs the agent loop, the messaging gateway, cron jobs and tools, and makes outbound HTTPS calls. That is why Nous describes Hermes as deployable on a “$5 VPS” as well as GPU clusters. A GPU instance would add cost and no speed for this setup.

The exception is local inference. If you want to run open-weight models yourself through Ollama or vLLM, you need a machine with enough GPU memory for the model, which is a different product category with a very different price tag. Even then the agent and the model server can live on different machines, so the cheap VPS still hosts Hermes while the GPU box serves the model. See Hermes Agent with Ollama and Hermes Agent on a cloud GPU.

VPS plans compared (October 2026)

Prices checked October 7, 2026 on each vendor's own pricing page or public pricing feed, in the currency the vendor shows. Hetzner's feed lists prices before VAT, and Contabo says its price includes applicable taxes; for the other vendors check the tax treatment at checkout. Cheap headline prices often require a long term, so check the “Term” column.

Two caveats on this table. First, Hetzner's cheaper CX and Arm CAX plans were showing as unavailable with no price on the cost-optimized page when we checked, so we left them out rather than quote an old number. Second, the long-term plans are cheaper per month because you commit up front. Hostinger's renewal rates are roughly double the introductory ones, and Contabo's price is an introductory rate for the first 24 months, and OVHcloud's assumes you pay for a year up front.

Our picks

• Cheapest that is still comfortable: OVHcloud VPS-1. Four gigabytes of RAM and two vCores for $4.54 a month is the lowest price we found with real headroom. You pay annually, so it is a commitment, not a trial.
• Best overall for most people: a month-to-month 2 vCPU / 4 GB plan from Hetzner, Vultr, DigitalOcean or Akamai. You pay more per month than the long-term deals, but you can resize or delete the server without losing a prepaid term, which matters while you are still learning what your agent needs. Vultr is the cheapest of the four at $20; Hetzner CPX22 (€19.49 before VAT) is similar once converted to dollars, and suits readers who want EU data centres. Per-provider walkthroughs: Hetzner, DigitalOcean, Hostinger. This is our judgment on price and flexibility, not a benchmark of CPU or network quality.
• Best free option: Oracle Cloud Always Free. Oracle's documentation lists Ampere A1 allocation of 1,500 OCPU hours and 9,000 GB hours a month, equivalent to 2 OCPUs and 12 GB running continuously, and Hermes supports aarch64 Linux. The catch is in the same document: instances are treated as idle and reclaimable if, over 7 days, CPU and network utilization stay under 20% (and memory under 20% on A1 shapes), and you can hit an “out of host capacity” error when creating one. A mostly idle agent can look exactly like an idle instance. Fine for experiments, risky for a bot you depend on.
• Best for US versus EU latency: pick the region nearest you, but do not over-think it. Hetzner lists EU (Falkenstein, Nuremberg, Helsinki) and US (Ashburn, Hillsboro) locations, and the US plans have different specs and prices, for example CPX11 at 2 vCPU / 2 GB and CPX21 at 3 vCPU / 4 GB. Vultr, DigitalOcean and Akamai offer many regions. Your agent's latency is dominated by the model provider's response time, so region matters mostly for dashboards, terminals and webhooks.

For a lower-cost overview that includes managed options, see cheapest Hermes Agent hosting.

Set up Hermes Agent on the VPS (Docker)

Hermes Agent ships as a Docker image and runs on any Linux VPS. The steps below use Ubuntu 22.04 or 24.04 LTS, and take roughly 30–60 minutes the first time.

Step 1: Install Docker

# On Ubuntu 22.04 / 24.04
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER
# Log out and back in for group changes to take effect

Verify Docker is working: docker run hello-world

Step 2: Pull the Hermes image

The official image is nousresearch/hermes-agent on Docker Hub. Upstream publishes latest and stable (the image that passed the stable release gate), main (development builds) and versioned X.Y.Z tags. Pin a versioned tag, or an image digest for an exact pin, so an update cannot change behavior under you:

# Pull a pinned stable version (see the upstream releases for the current one)
docker pull nousresearch/hermes-agent:X.Y.Z

# Verify the image is available
docker images | grep hermes-agent

Find the current recommended tag in the upstream Hermes repository.

Step 3: Create a persistent data volume

Hermes stores its configuration, session credentials and memory data at /opt/data inside the container. Mount a host directory there so the data survives restarts and updates:

# Create a directory on the host for Hermes data
mkdir -p /home/$USER/hermes-data

# The volume mount in your docker run command:
# -v /home/$USER/hermes-data:/opt/data

Without a persistent volume, Hermes configuration, WhatsApp/WeChat sessions and memory data are lost every time the container is replaced.

Step 4: Run the container

# Telegram, Discord and Slack connect outbound: no port needs publishing
docker run -d \
  --name hermes-agent \
  --restart unless-stopped \
  -v /home/$USER/hermes-data:/opt/data \
  -e OPENROUTER_API_KEY=sk-or-... \
  nousresearch/hermes-agent:X.Y.Z \
  gateway run

Replace X.Y.Z with the tag you pulled. The image's entrypoint runs hermes with the arguments you pass, so the command is just gateway run. The --restart unless-stopped policy brings the container back if the gateway exits or the server reboots. The chat platforms make outbound connections, so the container needs no published port for them.

The web dashboard is off by default. To turn it on, add -e HERMES_DASHBOARD=1 set a login with -e HERMES_DASHBOARD_BASIC_AUTH_USERNAME=... and -e HERMES_DASHBOARD_BASIC_AUTH_PASSWORD=..., and publish it on the host's loopback only with -p 127.0.0.1:9119:9119, then reach it through an SSH tunnel. The login is required: inside the container the dashboard binds to all interfaces, and on a non-loopback bind it refuses to start without an auth provider.

Step 5: Configure provider and model

# Interactive wizard: pick the provider and the default model
docker exec -it hermes-agent hermes model

Or edit /home/$USER/hermes-data/config.yaml directly and restart the container. Provider-specific guides: Anthropic, DeepSeek, Grok. For a non-Docker install, see the install guide.

Keeping it running 24/7

• Run it as a service. With Docker, --restart unless-stopped (above) brings the container back after a reboot. On a native install, Hermes can register its gateway with hermes gateway install; the docs recommend sudo hermes gateway install --system for a VPS so it starts at boot.
• Firewall everything except SSH. Enable UFW or your provider's firewall. A Hermes bot makes outbound calls and needs no open inbound ports unless you accept webhooks. If you do expose a webhook or the API server (port 8642), put a reverse proxy such as Caddy or nginx with HTTPS in front of it.
• Keep the dashboard private. The dashboard listens on port 9119, and upstream warns that opening any port on an internet-facing machine is a security risk. It requires authentication when bound to a non-loopback address. Restrict 9119 to your own IP, or prefer SSH tunnelling or a private network such as Tailscale.
• Patch and update. Enable unattended security updates on the OS. To update Hermes, pull the new image tag, stop and remove the container, and re-run the docker run command with the new tag; the data volume persists:

docker pull nousresearch/hermes-agent:NEW.X.Y.Z
docker stop hermes-agent
docker rm hermes-agent
# Re-run the docker run command with the new tag

• Back up ~/.hermes (native install) or your mounted data directory (Docker). That is where your memory, sessions and configuration live. A provider snapshot is not a substitute for a copy you can restore elsewhere.

Is the best VPS for OpenClaw the same?

Yes, almost entirely. OpenClaw is a Node.js project: its install documentation asks for Node 24.16+ or 26.1+ (Node 26 recommended) and its installer provisions Node if it is missing. Its VPS page lists Hetzner, DigitalOcean, Hostinger, Oracle Cloud and Raspberry Pi among supported hosts and recommends SSD-backed disks and a longer systemd startup timeout on slow machines. It publishes no minimum RAM figure, so we size it the way we size Hermes: 2 GB minimum, 4 GB for comfort. Details are in the OpenClaw VPS guide, and the managed option is on the hosting page.

VPS versus managed hosting

A VPS gives you root, any region, and the lowest monthly price on the long-term plans. It also gives you every job around the agent. The table is about where the hours go, not the invoice.

Self-hosting makes sense when you need data to stay in a specific jurisdiction or on hardware you control, want to reuse infrastructure you already pay for, need custom network routing, firewall rules or VPN integration, want to run Hermes next to a local Ollama or other private model server, or have requirements the managed platform does not cover.

OpenClaw Launch hosts Hermes Agent and OpenClaw. Lite has 1 vCPU, 2 GB RAM, 10 GB storage and 1 instance, with $1 a month of AI credits included. Pro has 2 vCPU, 4 GB RAM, 40 GB storage, up to 3 instances and $10 a month of AI credits. Lite lands on the same 2 GB size this guide calls comfortable for a chat-only bot, and Pro matches the 4 GB we recommend for browser use. Choose a VPS if you want root access, a specific region, or to run other services alongside the agent. Choose managed if you would rather spend the afternoon using the agent. More tradeoffs are in managed vs self-hosted and best Hermes Agent hosting.

Frequently Asked Questions

Related Guides

• Install Hermes Agent — full installation walkthrough
• Deploy Hermes Agent — deployment patterns for different environments
• Hermes Agent System Requirements — measured RAM, CPU and disk
• Best and Cheapest Hermes Agent Hosting — managed platforms ranked by total cost
• Hermes Agent on Hetzner
• Hermes Agent on DigitalOcean
• Hermes Agent on Hostinger VPS
• OpenClaw on a VPS — the same advice for OpenClaw
• Hermes Hosting and Hosting