No, and that's normal — that's not what the VPS does. It hosts the agent, the program that receives your instructions, calls a model, and does the work. The computation itself stays with the model provider (Anthropic, OpenAI, Mistral…), and you're billed per use with your own API keys. So no GPU is needed, just memory, a solid network, and a server that stays on. Running an LLM locally needs a machine with a GPU, and our VPS range doesn't offer that. If you still want a local model for small tasks (embeddings, lightweight classification), some compact models run on CPU with a few GB of RAM, but performance stays well below a commercial-grade LLM. For a production agent, the API remains the simplest choice: you switch provider or model with one line of configuration, without ever reinstalling a GPU driver or managing quantization.

No, it arrives bare (well, with the OS of your choice and full root access). Installing the AI agent runtime you want is on you. The server is ready in under 5 minutes, and you keep control of your versions and your data. Specifically, once your VPS is delivered, you connect over SSH, clone the repository for your chosen runtime (OpenClaw, Hermes Agent, a Python framework…), install its dependencies, then run it as a systemd service so it survives reboots. Expect roughly fifteen minutes to an hour for a first deployment, depending on the runtime. Our documentation covers installing the most requested runtimes, and technical support helps if you get stuck on the server side (network, firewall, resources) — installing and configuring the agent itself stays on your end.

Mostly on the runtime. As a rough guide, each project's official docs give ballpark figures: 1 GB RAM for Hermes Agent, 2 GB for OpenClaw or n8n (4 to 8 GB if OpenClaw drives a browser), 4 GB for Claude Code. CPU matters less than RAM here, since model computation goes to the provider over API. The simplest approach is to test with your real workload, then adjust vCore, RAM and storage from your customer area. Storage matters mainly if your agent keeps a long history or a local vector index: a few GB covers everyday use, more if you store embeddings or detailed logs. To get started, a 2 vCore / 4 GB RAM configuration covers most of the runtimes mentioned; you watch actual usage from your customer area and scale up as soon as RAM or CPU nears its limit, without migrating to another server.

Yes. Full root access, so you set up whichever provider you want (Anthropic, OpenAI, Mistral, open source, doesn't matter) with your own API keys. Nothing is blocked on external services, subject to our terms of use. You can even run several keys and providers side by side on the same agent: a fast, cheap model for sorting requests, a more capable one for complex tasks, each billed separately by usage. Keys are stored as ordinary environment variables (a .env file, a secrets manager): FirstHeberg has no access to them and never shares them. If a provider changes its pricing or rate limits, you switch to another one by simply editing your agent's configuration, without reinstalling anything on the server.

It depends on usage. You trade a flat fee for two variable costs, the VPS and model calls billed on actual use. What brings the bill down is that several agents can run on the same server and you can switch models depending on the task. For light, occasional use, a subscription stays simpler; for an agent running continuously, or several projects at once, self-hosting quickly pays off. A rough figure: a typical AI subscription runs around €20 to €30 a month for individual use, regardless of actual traffic. With a VPS at a few euros a month plus API calls billed per token, an agent handling a moderate volume often costs less once it runs several hours a day. The advantage disappears if the agent sits idle most of the time: in that case, a usage-based subscription stays easier to budget.

Yes, that's actually the whole point. Each agent is a process: you run them side by side, as separate systemd services or Docker containers, on the same machine. The only limit is available memory: not a license, not a seat count. Count each agent's RAM plus its dependencies (database, vector index), and bump resources from your customer area if needed. In practice, most users run between 2 and 5 agents on a mid-sized VPS (4 to 8 GB RAM) without any slowdown, as long as the agents don't all share the same local model or the same large database. A good habit is giving each agent its own system user and its own configuration folder: it avoids dependency conflicts and makes it easy to restart or stop a single agent without touching the others.

Yes, both ways. Nothing filters outbound calls to model APIs or your third-party services. To receive webhooks, the VPS has a public IPv4 — open the port you need and they arrive directly, no need for a temporary tunnel like you'd use from a home machine. The firewall stays under your control, like the rest of the system. For integrations that require an HTTPS URL (most modern webhook platforms do), you install a lightweight reverse proxy (Nginx or Caddy) with a free Let's Encrypt certificate in front of your agent: a few lines of configuration are enough, and certificate renewal happens automatically. You can also restrict webhooks to a list of allowed IPs or add a secret token to the URL to stop a third party from triggering your agent by mistake.

Run the agent as a systemd service with "systemctl enable" and it restarts on its own with the machine. On the infra side, we commit to a 4 business-hour repair time for hardware or network failures, with support reachable 24/7. Backups and snapshots stay available as an option from the customer area. To go further, a process manager like supervisor, or a simple Restart=on-failure rule in the systemd service file, restarts the agent automatically if it crashes, without waiting for a full server reboot. We monitor infrastructure availability (network, hypervisor, storage) 24/7 on our side; monitoring the agent's own application (logs, alerts on business errors) stays yours to set up — for example, with a script that notifies you on Telegram or by email after repeated failures, without waiting for you to notice it yourself.

No. The price shown is the monthly price, with no commitment and no increase on renewal: the same amount in month thirteen as in month one. You can increase vCore, RAM and storage whenever you want, and the price then simply follows the added resources. This is a policy applied across our whole VPS range, not a promotion limited to the AI agent offer: unlike some hosts that advertise an introductory price valid only for the first year, the price you see when ordering stays the price you pay afterward, month after month, with no surprise on the next invoice. If you change your configuration (more RAM to run an extra agent, for example), only the cost of the added resources is added to the bill; the base price never changes retroactively.