Self-host n8n on a Linux VPS

The reliable path on Linux is Docker Compose with n8n and PostgreSQL as two containers behind a reverse proxy for TLS (the same container stack as general Docker hosting):

1. Install Docker and the Compose plugin on Ubuntu, Debian or AlmaLinux with the distribution's package manager (for example apt install docker.io docker-compose-plugin).
2. Define two services in one docker-compose.yml: the n8nio/n8n image and a postgres image, each with a named volume so your data survives restarts and image updates.
3. Point n8n at PostgreSQL with DB_TYPE=postgresdb and the database host, name and credentials, so workflows and execution history are stored in Postgres instead of the default SQLite file.
4. Add a reverse proxy such as Caddy, Nginx or Traefik in front for HTTPS on your own domain, and set N8N_HOST and WEBHOOK_URL to that domain so webhooks resolve correctly.
5. Start it with docker compose up -d. Docker's own service is enabled under systemd, so the stack comes back on its own after a reboot.

Keep the Compose file and its environment variables under version control without the secrets, and schedule a regular backup of the Postgres volume.

Turn on n8n's owner account so the editor needs a sign-in, and where you can, reach the editor only over a VPN or from known IP addresses instead of exposing it to the whole internet. Publish only the webhook URLs that outside services genuinely need to reach. Each plan comes with a dedicated IPv4 address, so you can point your own domain at the server and terminate TLS at the reverse proxy. Automations must follow the terms and rate limits of the services and APIs they call; see the acceptable use policy.

n8n's AI and HTTP nodes send prompts to hosted models such as OpenAI and Anthropic using your own API key. The model runs on the provider's side and returns a response, so the heavy computation never touches your server: a CPU-only VPS is all n8n needs to orchestrate an agent, chain steps together and store the results. No GPU is involved.

If you want a model on the server itself, only a small quantized model through Ollama will run on the CPU, and it will be slow — useful for testing, not for production throughput. Training models and generating images need GPU hardware, which a VPS does not have.

Running agents this way is part of our wider AI VPS range: if you want a server set up specifically for the job, AI agent hosting covers long-running agents, and MCP server hosting covers the tool servers an agent reaches over the Model Context Protocol.