The moment I knew we had to move: an invoice arrived showing that our automation platform had charged us for 41,000 "tasks" in a month. About 30,000 of those tasks were a filter step deciding "no, ignore this one". We were paying, per unit, to be told nothing had happened.
(I printed that invoice and stuck it above my desk for a while. Marla made me take it down.)
What n8n actually is
It's a workflow automation tool — drag nodes onto a canvas, connect them, and things happen on a trigger. Same category as Zapier and Make. Two big differences: you can self-host it, and the pricing is per workflow execution, not per step. A workflow with forty steps costs the same as one with two.
It also has proper AI nodes now: an Agent node, vector store nodes, model connectors for OpenAI, Anthropic, Google, Ollama and others. That's what moved it from "handy plumbing" to "the backbone of our automation stack".
Our setup, exactly
- Host: a $12/month Hetzner VPS (4GB RAM, 2 vCPU). It has never been the bottleneck.
- Install: Docker Compose, with Postgres for the execution database and a Caddy reverse proxy for TLS.
- Backups: nightly Postgres dump to object storage. Workflows also exported as JSON into a private Git repo — this has saved us twice.
- Workflows: 34 live, ranging from "post to Slack when a deal closes" to Gwen's invoice-chasing agent (see that thread).
Pricing tiers
| Option | Cost | Notes |
|---|---|---|
| Community (self-hosted) | Free + your server (~$6–20/mo) | What we run. Fair-code licence. No SSO, no advanced permissions. |
| Cloud Starter | ~€24/month | 2.5k executions, 5 active workflows. Fine for a first project. |
| Cloud Pro | ~€60/month | 10k executions, 15 active workflows, better concurrency. |
| Enterprise / self-hosted paid | Custom | SSO, log streaming, environments, external secrets. Worth it above ~15 people. |
Check n8n's pricing page — the cloud tiers get revised fairly often.
The migration, honestly
Two days, mostly Saturday. Roughly:
- Listed all 41 existing automations. Deleted 7 nobody could explain. (Do this step. It's the best hour of the whole project.)
- Rebuilt the 12 highest-value ones first, ran them in parallel with the old platform for a week.
- Rebuilt the rest, mostly copy-paste patterns by then.
- Spent an unexpected three hours discovering that self-hosted webhook URLs need your public domain configured properly, or nothing external can reach you. The docs say this. I did not read that bit.
- Cancelled the old subscription. Felt magnificent.
What we loved
- Predictable cost. The bill doesn't grow when a workflow gets more complicated. Enormous psychological difference — we stopped optimising for the pricing model.
- Code nodes when you need them. Drop into JavaScript or Python for the awkward 5% instead of building a Rube Goldberg machine of no-code steps.
- The AI Agent node is genuinely good. Tools, memory, model choice, all visual. Non-developers on our team have built working agents with it.
- Data stays on our box. Solved three separate conversations with clients about where their data goes.
- Excellent error handling. Per-workflow error branches and a global error workflow. Ours posts to Slack with a link straight to the failed execution.
What we'd warn you about
- You now run a server. Updates, TLS renewals, disk filling up with execution logs. Budget an hour a month, and actually spend it.
- Fewer polished integrations than Zapier for niche SaaS. You'll write HTTP request nodes. Not hard, but it's work.
- The canvas gets messy fast. Past about 25 nodes, readability collapses. Use sub-workflows earlier than feels necessary.
- Fair-code licence, not classic open source. Read it if you're planning to resell automation to clients — there are restrictions.
- Version upgrades occasionally break a node. Pin your image version, upgrade deliberately, keep those Git-exported JSON backups.
The money, 26 months on
| Old per-task platform | n8n self-hosted | |
|---|---|---|
| Monthly platform cost | $652 | $12 (VPS) |
| Extra maintenance time | ~0 hrs | ~1 hr/month |
| Workflows we avoided building because of cost | "loads" | None |
| 26-month total | ~$16,950 | ~$312 + ~26 hrs |
Crew verdict — 9.1 / 10
If you have someone comfortable with Docker and a domain name, this is the highest-leverage swap available to a small business right now. The predictable pricing changes behaviour: you start automating things that were previously "not worth a task credit".
Skip it if: nobody on the team wants to own a server, or you need SOC 2 paperwork from your automation vendor tomorrow. In that case take n8n Cloud, or stay on Zapier and accept the bill.
Official links
- n8n homepage
- Self-hosting documentation
- AI & agent node docs
- Community forum — genuinely responsive
Crew replies
I'm the non-technical proof point here. I built my invoice workflow with about four hours of Tobi's help and I've maintained it myself since. The visual canvas is the reason. I would not have done it in code.
Second the sub-workflow advice, hard. We had one monster workflow with 61 nodes. Splitting it into five named sub-workflows took a day and made the thing maintainable by someone other than its author.
Next up
If you're picking your automation backbone, read this alongside the model teardown — they solve different halves of the same problem.
Claude Sonnet 4.5 → All teardowns