How to self-host a private AI on Contabo
Quietly, Contabo has built a whole 1-click AI lineup — a local LLM chatbot you can run like a private ChatGPT, and a set of AI agents like OpenClaw. It's genuinely one-click. The part nobody tells you honestly is what actually runs well on a cheap CPU VPS versus what needs a GPU. Here's the real picture.
Verified against Contabo's help docs, Aug 2026. Only one 1-click app runs per server.
Contabo's AI 1-click apps
Two flavours, and the difference matters for what hardware you need:
Ollama Phi4 — a local LLM chatbot
Runs the model on your server: Ubuntu 24.04 + Ollama + a web chat UI + Microsoft's Phi-4 model, served over HTTPS. This is the "private ChatGPT" one — and the one that actually needs real RAM, because the model runs locally.
OpenClaw, ZeroClaw, Paperclip AI, Hermes Agent — AI agents
These orchestrate AI rather than run a model. You supply your own OpenAI-style API key; the agent calls that external LLM and handles tasks, memory and integrations. Because the model runs at the API, these are happy on a cheap CPU VPS.
Path 1 — a private LLM chatbot (Ollama Phi4)
Select Ollama Phi4 as the 1-click app when ordering a VPS (or add it to an existing one via the control panel), set a password, and in ~30 minutes you get a browser chat UI over a local model at a Contabo HTTPS subdomain. No Docker, no reverse proxy to wire up.
The RAM reality (this is the honest part)
It's CPU-only — no GPU. The full Phi-4 (14B) wants 16 GB RAM minimum (24 GB recommended), i.e. a Cloud VPS 30-class box. On a common 8 GB VPS you run smaller models instead:
| Model | Fits 8 GB? | CPU speed (rough) |
|---|---|---|
| phi-4-mini / Llama 3.2 3B | Yes, comfortably | ~15–25 tokens/sec — responsive in a chat UI |
| Qwen / Gemma 4B | Tight but workable | ~10–20 tokens/sec |
| A 7B model | Only just | ~5–10 tokens/sec — sluggish |
| Phi-4 full (14B) | No — needs 16 GB+ | Low single digits on CPU |
Bottom line: for a single user chatting with a small model, a CPU VPS is genuinely usable — pick phi-4-mini or a 3B on 8 GB, or size up to a 16–24 GB Contabo plan for the full Phi-4. Ollama can pull other models too (Llama 3.2, Qwen, DeepSeek) from the command line.
Path 2 — AI agents (OpenClaw and friends)
If you don't need the model on your box — just an agent that uses one — the agent 1-click apps are the cheaper play. OpenClaw is an open-source AI agent for task automation and messaging integrations; you give it your own OpenAI-style API key and it does the orchestrating. ZeroClaw (lightweight autonomous runtime), Paperclip AI (multi-agent orchestration) and Hermes Agent (persistent memory) are variations on the same idea.
Because the LLM runs at the API provider, these need almost no local horsepower — a base CPU VPS is fine, and your only model cost is the API usage. This pairs naturally with workflow automation: an agent for reasoning, plus something like self-hosted n8n for the deterministic steps around it.
CPU VPS or GPU Cloud?
Stay on a cheap CPU VPS if you're a single user running small local models, or running agent apps that call an external API. Move to Contabo's GPU Cloud (from around €690/mo) only when you want models above ~7B at interactive speed, reliably high throughput, long contexts, or several concurrent users. For most people experimenting with a private AI, the CPU box plus a small model — or an agent on an external API — is the sensible, cheap starting point.
Frequently asked
Can I run a private ChatGPT on a cheap Contabo VPS?
Yes, with caveats. Contabo's Ollama Phi4 1-click image installs Ollama, a web chat UI and a model with HTTPS ready to go. But it's CPU-only: the full Phi-4 (14B) wants 16 GB RAM, so on an 8 GB box you should run a small model — phi-4-mini or a 3B like Llama 3.2 — at roughly 15–25 tokens/sec. That's genuinely usable for a single user and small models; it is not a fast, big-model setup.
What is OpenClaw on Contabo?
OpenClaw is one of Contabo's free 1-click apps — an open-source AI agent for task automation and messaging integrations. It doesn't run a model itself; you plug in your own OpenAI-style API key, and the agent orchestrates that external LLM. Because the heavy lifting happens at the API, OpenClaw runs fine on a cheap CPU VPS.
Do I need a GPU?
Only for bigger models or real speed. A CPU VPS is fine for a single user chatting with a 3B/mini model, or for agent apps (OpenClaw, ZeroClaw, etc.) that call an external LLM API. You need Contabo's GPU Cloud once you want models above ~7B at interactive speed, higher throughput, long contexts, or multiple concurrent users — but that's a far pricier product (from around €690/mo).
Which of Contabo's 1-click apps are AI?
The AI-relevant ones are Ollama Phi4 (a local LLM chatbot) plus a set of agent frameworks: OpenClaw, ZeroClaw, Paperclip AI and Hermes Agent. n8n is there too for workflow automation. Note you can only run one 1-click app per server.
Related
Some links here are affiliate links — we earn a commission on Contabo, at no extra cost to you, and it never changes the recommendation. See our Affiliate Disclosure.