Server for AI agents

An agent that only runs while your laptop is open is not really running. This is a machine that stays up, in the EU, with your keys on it and nobody else on it.

Tier III Certified ISO 9001 ISO/IEC 27001 ISO 14001 GDPR Compliant

Who it is for

Someone running a personal assistant agent

It is connected to your messengers and your calendar, and it is offline every time you shut the lid or the wifi drops.

A team automating work with agents

The automation runs on somebody laptop, nobody else can reach it, and when that person is on holiday it stops.

Anyone who does not want prompts leaving Europe

The work involves customer data, and sending every prompt to an API outside the EU is not something legal will sign off.

What you run on it

  • OpenClaw, the personal assistant agent, staying online and reachable through your messengers instead of only while a laptop is open
  • Hermes, the open-weights model family, served locally through vLLM, Ollama or llama.cpp so prompts never leave the machine
  • Other open-weights models on the same footing: Llama, Qwen, Mistral, DeepSeek, GLM
  • Agent frameworks and schedulers: n8n, LangGraph, CrewAI, or a plain cron job that calls your own script
  • A private inference endpoint on your own network, so your other services talk to it instead of to a third-party API
  • Vector storage and the agent memory alongside it, on the same disk, backed up nightly

Sized by what the agent has to hold in memory

An agent that calls an external model is small: it is mostly waiting on the network. Serving a model yourself is the opposite, and memory is what decides which one fits. A quantised 8B model runs in these sizes on CPU; anything larger, or anything that has to be fast, wants a GPU, which we quote separately from /gpu-server.

One agent, external model
15
per month

Order Now
Assistant calling an API, always on
2 vCPU
4 GB RAM
40 GB NVMe
1 static IPv4
1 backup included
Agent with a browser and memory
18.43
per month

Order Now
Web access, vector store, several tools
4 vCPU
8 GB RAM
80 GB NVMe
1 static IPv4
1 backup included
Self-hosted model on CPU
59.92
per month

Order Now
Quantised 8B served locally, nothing leaves the box
8 vCPU
32 GB RAM
160 GB NVMe
1 static IPv4
1 backup included

What is and is not included

  • Model weights and API keys are not included. Open-weights models are yours to download under their own licences, and a commercial API stays on your account.
  • GPU acceleration is not part of these prices. CPU inference is fine for a quantised small model and slow for anything else, so ask and we will quote the card that fits.
  • This is not a managed agent platform. It is a server with root access: you choose the runtime, the model and the update schedule.
  • An agent is not an exemption from the acceptable use policy. Bulk messaging, scraping at scale and credential testing are out, whoever or whatever initiates them.
  • We do not inspect what your agent does. That also means we cannot debug it for you unless you buy server administration.

Running an agent that nobody is watching

  • Always on is the whole point. A personal agent is judged on whether it answered while you were asleep, and that is a hosting question rather than a model question.
  • Give it its own machine. An agent holds messenger sessions, API keys and often write access, and none of that belongs next to your production database.
  • Snapshot before every upgrade. Agent projects move fast and break their own state; a rollback in two clicks is cheaper than reconstructing a memory store.
  • Watch the egress. An agent that browses from a dirty address gets captchas and blocks, so ours are clean and you can set rDNS on them.
  • Cap the spend. An agent in a retry loop against a paid API is the most expensive bug in this field, and the limit belongs on the key, not in the prompt.
  • Keeping the model local is what makes the GDPR conversation short: the prompt, the context and the output never leave a Tier III facility in the EU.

A hosted agent service versus your own server

Hosted agent serviceYour own DCXV server
Where the prompts goThrough the provider, wherever it isNowhere, if you serve the model yourself
Which agent you may runThe ones the platform supportsAnything that runs on Linux
Version and update timingWhen the platform decidesWhen you decide
Keys and messenger sessionsStored by the providerOn your disk, under your root
Recovering from a bad upgradeOpen a ticketSnapshot rollback in two clicks
CostPer seat, per run, or per tokenFrom EUR 15/mo, monthly, no lock-in

Getting set up

1

Say what the agent does

Whether it calls an external model or serves one itself, and whether it needs a browser. That is enough to pick a size.

2

We hand over the server

A cloud VM in Prague or Covilha with a static address and root access, deployed in minutes. Boot your own ISO if you would rather.

3

Install the agent and pin it

Its own user, its own directory, secrets outside the repository, and a spending cap on any paid key before it runs unattended.

4

Snapshot, then leave it running

Take the snapshot once it works. Every upgrade after that is a two-click rollback instead of an evening.

Why choose us

  • Tier III certified facilities, 99.982% facility SLA
  • Own network, AS204057, IPv4 and IPv6 dual-stack
  • 24/7/365 support with ~10 minute average response
  • Cyprus company, EU jurisdiction, GDPR-native since 2007

FAQ

- How much does a server for an AI agent cost?

From EUR 15 per month for an agent that calls an external model: 2 vCPU, 4 GB RAM, 40 GB NVMe and a static IPv4. Add a browser and a vector store and it is EUR 18.43 for 8 GB. Serving a quantised 8B model on the box itself is EUR 59.92 for 8 vCPU and 32 GB. GPU acceleration is quoted separately

- Can I run OpenClaw on it?

Yes, and an always-on server is what the agent is missing on a laptop: it stays reachable through your messengers while you are asleep instead of only while the lid is open. It runs as its own user with its own directory, you hold the keys, and a snapshot before each upgrade means a bad release is a rollback rather than a rebuilt memory store

- Can I self-host Hermes or another open-weights model?

Yes. Hermes, Llama, Qwen, Mistral, DeepSeek and GLM all serve locally through vLLM, Ollama or llama.cpp. A quantised 8B model is usable on CPU in our 32 GB size; anything larger, or anything that has to answer quickly, wants a GPU, and we have NVIDIA L4, L40S, A40, A100 and H100 available. The weights are yours to download under their own licences

- Do my prompts leave the EU?

Not if you serve the model yourself. The prompt, the context and the output stay on a machine in a Tier III facility in Prague or Covilha, which is what makes the compliance conversation short. If your agent calls a third-party API instead, that traffic goes wherever that provider is, and that is worth checking before customer data is involved

- Will my agent get blocked or captcha-flagged when it browses?

That depends on the reputation of the address it browses from, which is the part most hosts ignore. We are an approved RIPE NCC broker and run our own space under AS204057, so the addresses are clean and you can set rDNS on yours. It is the one advantage here that a reseller cannot copy

- Can I use it for scraping or bulk messaging?

No. An agent doing it is the same as a person doing it: scraping at scale, credential testing and bulk unsolicited messaging are outside our acceptable use policy. We also keep agent workloads on separate address ranges from the space that goes to mail projects and to sale, so one abusive tenant cannot spoil an asset other customers depend on

If you require assistance or have additional questions, please contact the managers or write to the support team at support@dcxv.com

Cloud servers from €15/mo

Order Now