What Was Tested
Self-hosted n8n via npx, with a 3-node AI workflow: Manual Trigger → HTTP Request → Ollama.
Environment:
- macOS Apple Silicon
- Node.js v22.15.0 / npm 10.9.2
- repoverifiertest user (clean environment)
- No Docker — npx only
- ollama running on host with llama3.1:8b pulled
Install
npx n8n
Single command. Downloads and starts n8n in under 2 minutes.
No config files. No Docker. No environment variables required.
Editor accessible at http://localhost:5678 immediately.
What Worked
Self-hosting:
npx n8n ran cleanly on first attempt. Setup wizard created a local
account in under 60 seconds. Visual editor loaded immediately.
HTTP Request node:
Added as node 2. Set URL to https://news.ycombinator.com.
Clicked Execute step — raw HTML returned on first attempt.
Zero configuration beyond the URL field.
Native Ollama node:
n8n ships with a built-in Ollama node — no community packages needed.
Two actions available: Analyze image and Message a model.
Selected Message a model, created credential with Base URL
http://127.0.0.1:11434, model list auto-populated with llama3.1:8b.
Data flow between nodes:
Raw HTML from HTTP Request node passed automatically to Ollama node
via {{ $json.data }} — no manual wiring or transformation needed.
AI summarisation:
Prompt: "Summarise the top 5 stories from this Hacker News page in
plain English: {{ $json.data }}"
llama3.1:8b returned a structured summary of HN stories correctly
extracted from raw HTML. Output clean and readable on first run.
Full workflow execution:
Manual Trigger → HTTP Request → Ollama ran end to end without errors.
Total time from npx n8n to working AI workflow: under 30 minutes.
Honest Findings
localhost fails for Ollama credential — use 127.0.0.1:
When creating the Ollama credential, http://localhost:11434 returns
"Couldn't connect with these settings — the service refused the
connection." Changing to http://127.0.0.1:11434 connects immediately.
This is not documented in the README. Costs 2 minutes if you don't know.
Raw HTML passed to LLM without preprocessing:
The workflow passes raw HTML directly to llama3.1:8b. This works but
is inefficient — llama3.1:8b processes boilerplate and markup alongside
content. For production use, add a trafilatura step between HTTP Request
and Ollama to extract clean text first.
No persistent storage by default:
npx n8n stores data in ~/.n8n on the local filesystem. Workflows and
credentials are preserved between restarts but not production-hardened.
For persistent production deployment use Docker with a mounted volume
or the official n8n-hosting configurations.
What Was Not Tested
- Scheduled triggers (cron-based automations)
- 400+ integration nodes beyond HTTP Request and Ollama
- Multi-step AI agent workflows with memory
- Production deployment with PostgreSQL backend
- Cloud deployment via n8n-io/n8n-hosting