n8n is already one of the most-deployed self-hosted tools out there - it's the open-source alternative to Zapier/Make, and it's usually the first thing people put on a fresh VPS. What's changed recently is what people are wiring it up to: instead of "when a form is submitted, send a Slack message," a lot of n8n workflows now have an LLM call sitting in the middle of them - summarize this ticket, draft this reply, classify this lead.
The problem is that the moment you add an "OpenAI node" to a workflow that's processing customer data, you've sent that data to a third party, on every run, forever. Pointing the same workflow at a model you run yourself removes that entirely.
The stack
- n8n - the workflow engine: triggers, HTTP requests, conditionals, the automation logic.
- Ollama - the model n8n calls instead of an external API.
n8n has a built-in "Ollama" node (and a generic HTTP Request node works too, since Ollama exposes a normal REST API) so swapping an OpenAI-node workflow over to a self-hosted model is usually a five-minute change, not a rebuild.
Deploying it
- From Templates, deploy n8n and Ollama on the same server - either individually or, if you also want a chat interface for testing prompts outside of a workflow, deploy the AI Workbench stack (Ollama + Open WebUI + Flowise) alongside n8n.
- Both containers land on the same private Docker network, so n8n reaches Ollama at
http://ollama:11434with no manual networking - the same wiring DeployOS uses to connect Open WebUI to Ollama. - In n8n, add an Ollama node (or an HTTP Request node pointed at that address) anywhere you'd have used an OpenAI node.
Why this is also a Coolify-migration story
If you're already running n8n on Coolify or a hand-rolled Docker Compose setup, moving it doesn't mean rebuilding your workflows - DeployOS's discovery flow scans a server for running Compose projects and can adopt an existing n8n instance as a managed app without tearing it down first. You get automatic HTTPS renewal, health-checked deploys, and backups on top of a workflow library you already built, plus a place to add Ollama next to it without standing up a second server.
What this actually buys you
Not a smarter model than GPT-4 - a private one. The workflows that make sense to automate this way are the ones where the input is something you wouldn't want an OpenAI usage log for: internal documents, customer PII, anything under an NDA. For everything else, keep using whatever API node is fastest. The point of self-hosting the model isn't ideology, it's having the option to route the sensitive third of your workflows somewhere that isn't someone else's server.