My Self-Hosted AI Stack: Everything Running on One €4/Month Server
I run 23 active automation workflows, 7 content websites, an AI proxy, a newsletter platform, and a vector database — all on a single server that costs €3.99 per month.
This is the exact setup, why each piece is there, and what I’ve learned maintaining it.
Why Self-Host at All?
The honest answer: cost and control.
The managed equivalent of this stack would run €200–500/month. Zapier for automation. OpenAI API with no cost cap. Mailchimp for newsletters. Cloudflare for storage. Hosted analytics.
Self-hosting the whole thing costs under €20/month. The trade-off is time spent on maintenance — but with Coolify managing containers, that’s closer to 30 minutes a month than a full-time job.
The second reason is no vendor lock-in. When a provider raises prices or changes an API, I swap it in LiteLLM and nothing else changes.
The Server
Hetzner CAX11 — ARM architecture, 2 vCPU, 4GB RAM, 40GB SSD, 20TB traffic, €3.99/month.
The ARM architecture matters. Hetzner’s ARM servers have significantly better performance-per-euro than their x86 equivalents. I was skeptical at first — most Docker images are x86 — but multi-arch support has become good enough that I’ve had no compatibility issues with any service in this stack.
Located in Helsinki. Latency to Stockholm is ~15ms. Not relevant for async workflows, but nice for the occasional direct API call.
4GB RAM sounds tight. Here’s actual usage with everything running:
n8n: ~280MB
LiteLLM: ~180MB
Minio: ~120MB
Umami: ~90MB
Postgres (x2):~160MB each
Listmonk: ~70MB
Qdrant: ~150MB
Coolify: ~200MB
Total: ~1.4GBLeaves about 2.4GB free. The server has never OOM-killed a container.
Coolify: The Foundation
Coolify is what makes running 12+ containers manageable without becoming a full-time sysadmin.
What it handles:
- Container lifecycle: deploy, restart, update with one click
- HTTPS: Traefik auto-issues Let’s Encrypt certs for every custom domain
- Environment variables: stored per-service, never in code
- One-click services: Postgres, Redis, Minio, n8n are all available as pre-configured templates
Before Coolify I managed Docker Compose files manually. Every new service meant writing a compose file, handling the Traefik labels for HTTPS, managing env files. Coolify wraps all of that.
The one thing Coolify doesn’t do: it won’t tell you when a container is unhealthy without you checking. I solved that with a daily Telegram report (more on that below).
n8n: The Orchestration Layer
n8n is where the logic lives. Every automated action in the stack is an n8n workflow:
- Daily keyword research for each site (runs at 08:00–09:45 depending on site)
- Article generation and publishing (10:00–12:00)
- Analytics reports (Umami stats to Telegram every morning)
- GEO/AEO intelligence (daily RSS parsing + Groq analysis)
- AI traffic monitoring (daily Umami Postgres query)
- Monthly performance report (1st of month, 09:30)
- Error monitoring (Telegram alert on any workflow failure)
23 active workflows total. n8n runs as a container with a Postgres backend for execution history.
What n8n is good at: chaining API calls, conditional logic, scheduled tasks, webhook endpoints. The visual editor makes it easy to debug — you can see exactly what data each node received and what it output.
What n8n is not good at: complex data transformations (use a Code node), fan-out patterns where multiple branches merge back (they break silently on empty results — use sequential chains instead).
The most important n8n lesson I learned: use sequential chains. Connect nodes in a single line A → B → C → D. Never have two nodes both connect to a third without a Merge node in between — empty results from one branch silently kill the chain.
LiteLLM: The AI Gateway
Every LLM call in the entire stack goes through LiteLLM. n8n workflows call http://litellm/v1/chat/completions, not the OpenAI API directly.
Why this matters:
Cost visibility. LiteLLM tracks spend per request, per model, and per day. I can see exactly what each workflow costs. My total March spend was $3.76 across all 23 workflows. Without LiteLLM I’d have no idea which workflow was expensive.
Provider flexibility. My article generators use Groq (fast, cheap, llama-3.3-70b). My hockey sites also use Gemini (free tier). LiteLLM routes to the right provider based on the model name. If Groq raises prices tomorrow, I change one config line — nothing in the workflows changes.
Rate limit handling. Groq has generous rate limits but they exist. LiteLLM queues requests and handles retries. Workflows never see a rate limit error.
The model I use most: llama-3.3-70b-versatile via Groq. It’s fast (~1500 tokens/sec), good at following formatting instructions, and costs roughly $0.05 per 2000-token article output.
Minio: Workflow-to-Workflow Data Passing
This one surprised me. I didn’t expect object storage to become so central to the architecture.
The problem it solves: n8n workflows don’t share state. If keyword research runs at 09:45 and the article generator runs at 10:00, they need a handoff mechanism. You could use a database, but that’s overkill for a JSON file.
Minio is a self-hosted S3-compatible object store. I use it as a file system for workflow data:
seo-drafts/
├── tha-approved-keywords.json (written 09:45, read 10:00)
├── eik-approved-keywords.json (written 08:00, read 09:00)
├── performance-insights.json (written Mon 09:15, read by keyword research)
├── geo-aeo-insights.json (written daily 07:00, read by monthly report)
├── ai-traffic-2026-03-30.json (daily AI traffic snapshot)
└── [article drafts waiting for approval]n8n has a native S3 node that reads and writes these files. No custom code needed.
Minio also stores EIK article drafts while they wait for my manual approval. The draft sits in Minio, I get a Telegram message with Approve/Reject buttons, and if I approve, the workflow reads the draft from Minio and pushes it to GitHub.
Umami: Analytics Without the Tracking Overhead
Umami is a self-hosted analytics platform. It runs on Postgres and tracks pageviews, sessions, and custom events — no cookies required, GDPR-compliant by default.
Every site in the stack has the Umami script installed. I query the Umami Postgres database directly from n8n to build daily and weekly analytics reports.
One specific use I’m pleased with: tracking AI referrer traffic. I filter sessions where the referrer contains perplexity.ai, chatgpt.com, claude.ai, or copilot.microsoft.com. This lets me measure whether my GEO (Generative Engine Optimisation) work is having any effect. It’s early data but the trend is upward.
The Monitoring Layer
A self-hosted stack with no monitoring is a stack that breaks silently. I have two layers:
Daily infra report (08:05, Telegram):
- Server uptime, load average, RAM usage, swap, disk
- Status of every key container (✅ healthy / 🟡 up-no-healthcheck / ❌ down)
- Generated by a bash script via cron, no n8n dependency
n8n error monitor:
- Any workflow execution that ends in an error triggers an immediate Telegram notification
- Includes the workflow name, node that failed, and error message
- Built as an n8n workflow listening to the n8n error event
These two together mean I know about any problem within minutes, usually before it affects anything.
The Full Cost Breakdown
| Service | Cost |
|---|---|
| Hetzner CAX11 | €3.99/month |
| Domain registrations (7 domains) | ~€5/month avg |
| Groq API (LLM inference) | ~$3–5/month |
| Netlify (hosting, free tier) | €0 |
| GitHub (private repos) | €0 |
| All self-hosted services | €0 |
| Total | ~€15/month |
For comparison: Zapier’s Business plan starts at $49/month and doesn’t include AI, analytics, storage, or hosting.
What I Would Change
More RAM. 4GB works but leaves no buffer. If I add another service, I’ll either need to optimize or upgrade to the CAX21 (8GB, €7.49/month). Still cheap.
Backups. I have Coolify’s backup feature configured for Postgres, but I haven’t tested a restore. This is on the list.
Staging environment. Right now changes go straight to production. A failed workflow update can break a live pipeline. A second €4 server as a staging environment is worth it.
How to Replicate This
If you want to build a similar stack:
- Provision a Hetzner CAX11, point a domain at it
- Install Coolify following their one-line installer
- Deploy n8n, Postgres, and Minio as Coolify services
- Deploy LiteLLM with your API keys configured
- Start building workflows
The whole initial setup takes about 4 hours. Getting the workflows stable took considerably longer — that’s where the actual complexity lives.
If you’re building something like this for a business context — internal automation, content operations, data pipelines — that’s exactly the kind of project I help with. The architecture decisions and the pitfalls are non-obvious until you’ve hit them.
Reach out if it’s relevant.
Running costs are as of March 2026. Hetzner pricing may vary by region.