Self-hosted n8n costs 7.09 EUR a month in server bills on the smallest sensible production box. That is a Hetzner CX23 at 5.49 EUR excluding VAT for new orders since 15 June 2026 (Hetzner), automated backups at a flat 20% of the server price (Hetzner) and 0.50 EUR for a primary IPv4 address (Hetzner). Prices retrieved 26 September 2026. The total is our arithmetic.
The invoice is the cheap part. n8n ships with defaults that fail without telling anyone: full execution data kept on every success, a health check that ignores the database, schedules on New York time. Each has a documented fix. This guide pairs every failure with its setting.
If you are still choosing a tool, the Zapier vs n8n vs Make vs custom code comparison covers that decision. The broader argument for running your own infrastructure is in Own Your Stack.
By Dan Colta. NodeSparks is a two-person EU automation studio building owned automations for SME teams. We do not sell n8n hosting. Every price links to the vendor's own page.
Key Takeaways
- A single-box production n8n on a Hetzner CX23 costs 7.09 EUR a month excluding VAT, by our arithmetic on Hetzner prices. Queue mode on a CX33 is 10.69 EUR.
- Most silent failures are shipped defaults: execution data saved in full (n8n docs), schedules on New York time (n8n docs) and a /healthz check that ignores the database (n8n docs).
- An Error Trigger fires only when an execution fails. A heartbeat monitor such as Healthchecks.io, free for 20 checks (Healthchecks.io), catches the workflow that never ran.
- n8n asks self-hosters to update at least once a month (n8n docs). Censys counted 103,476 potentially vulnerable instances after a critical n8n flaw in December 2025 (Censys).
What does self-hosted n8n cost a month in production?
7.09 EUR a month excluding VAT for a single-box install on a Hetzner CX23 with Postgres in Docker. The server is 5.49 EUR for new orders since 15 June 2026 (Hetzner). Backups add a flat 20% (Hetzner) and a primary IPv4 adds 0.50 EUR (Hetzner). Queue mode on a CX33 comes to 10.69 EUR.
| Line item | Input | CX23, single box | CX33, queue mode | CX43 |
|---|---|---|---|---|
| Server | New orders since 15 June 2026 | 5.49 | 8.49 | 15.99 |
| Automated backups | 20% of the server price | 1.10 | 1.70 | 3.20 |
| Primary IPv4 | Flat monthly charge | 0.50 | 0.50 | 0.50 |
| Monthly total | Our arithmetic | 7.09 | 10.69 | 19.69 |
| Yearly total | Monthly x 12 | 85.08 | 128.28 | 236.28 |
EUR excluding VAT, Germany and Finland, retrieved 26 September 2026. A CX23 is 2 vCPU, 4 GB RAM and 40 GB NVMe (Hetzner).
The box is ordinary. We self-host Postiz, the open-source social media scheduler, on a Hetzner CX23 in the EU, live since 27 August 2026 (our Buffer alternative write-up). We do not run n8n on it. Every n8n default below comes from n8n's documentation rather than from our logs.
Hetzner's 15 June 2026 adjustment raised all three sizes by 30.8% to 37.6% (our arithmetic on Hetzner's table). Cost figures published before June use the old row.
Two optional lines bill in US dollars and sit outside the totals. Hetzner's backups live at the same provider as the server, which makes them a convenience rather than an off-site copy. Backblaze B2 stores the first 10 GB free, then charges $6.95 per TB a month (Backblaze). Heartbeat monitoring on Healthchecks.io is free for 20 checks and $20 a month for 100 (Healthchecks.io).
n8n's entry paid self-hosted plan, Business, is 667 EUR a month billed annually for 40,000 executions, servers extra (n8n). That is 94 times the CX23 bill (our arithmetic). n8n also asks self-hosters to "try to update at least once a month" (n8n docs): 12 maintenance windows a year to price at your own hourly rate.
What does the free n8n Community edition leave out?
Log streaming, external secrets, Git version control, environments and multi-main mode, among other features (n8n docs). The free edition runs under the Sustainable Use License, which permits use for your own internal business purposes (n8n on GitHub). Paid self-hosted features start at 667 EUR a month (n8n).
The same page lists external storage for binary data, projects, custom variables, SSO (SAML and LDAP) and sharing. Free registration unlocks folders, debug in editor and custom execution data.
External secrets is the gap that bites first. Without it, credentials stay in n8n's database, encrypted with a key n8n generates on first launch and saves in ~/.n8n (n8n docs). Lose that key and a restored database "can't be decrypted" (n8n docs). Set N8N_ENCRYPTION_KEY yourself and store it off the box.
The licence allows distribution to others only free of charge for non-commercial purposes (n8n on GitHub). Read it before building a hosted offer on n8n. If the paid line is where n8n stops fitting, n8n alternatives and when to skip n8n prices the exits.
Why does self-hosted n8n fail silently?
Because its defaults assume someone is watching. Failure alerts are set up one workflow at a time (n8n docs). EXECUTIONS_TIMEOUT defaults to -1, which means no timeout (n8n docs). The basic /healthz check does not look at the database (n8n docs). Each default has a documented fix, listed below.
| Silent failure | Shipped default | Fix |
|---|---|---|
| Execution data fills the disk | Pruning at 336 hours or 10,000 executions | Tune EXECUTIONS_DATA_MAX_AGE and EXECUTIONS_DATA_PRUNE_MAX_COUNT |
| Personal data in execution logs | Success and error saving both all; redaction Enterprise-only | Turn success saving off per workflow |
| A failed run alerts nobody | Error workflow chosen per workflow | Error Trigger workflow on every production workflow |
| A workflow never runs | Nothing alerts on an absence | Heartbeat ping at the end of each scheduled workflow |
| Google credential dies after a week | 7-day tokens for External apps in Testing | Publish the OAuth app |
| Webhook caller cut off | nginx proxy_read_timeout 60s | Raise it, or respond before the work finishes |
| A run never ends | EXECUTIONS_TIMEOUT -1 | Set an instance-wide timeout |
| Schedules at the wrong hour | GENERIC_TIMEZONE America/New_York | Set your zone |
| Queue worker dies mid-job | No stalled-job retry since 2.0; worker health endpoint off | QUEUE_HEALTH_CHECK_ACTIVE=true plus an external check |
| Database down, /healthz green | /healthz ignores the database | Monitor /healthz/readiness |
| Credentials will not decrypt after restore | Encryption key auto-generated in ~/.n8n | Set N8N_ENCRYPTION_KEY and back it up |
Sources: n8n's documentation, Google's OAuth 2.0 guide and nginx, retrieved 26 September 2026 and linked in the sections below.
How do you keep n8n execution data from holding personal data?
Save less of it. n8n keeps the full data of every successful and every failed production execution, because both save settings default to "all" (n8n docs). Pruning deletes it after 336 hours or once the total passes 10,000 executions (n8n docs). Until then, every email address a workflow touched sits in the database.
The paid fix is redaction, which hides execution input and output. It is Enterprise-only on n8n Cloud and self-hosted alike, with "Do not redact" as the default (n8n docs).
On the Community edition, turn success saving off for every workflow that handles customer personal data. Shorten EXECUTIONS_DATA_MAX_AGE from its 14-day default if nobody reviews old runs. For an EU company, two weeks of customer data that nobody reads is a retention decision made by default, the same question behind running AI agents unattended.
Even with pruning on, n8n warns that the database "can grow in size and run out of storage" (n8n docs). Annotated executions are never pruned. SQLite, the default database (n8n docs), reuses pruned space without releasing it to disk. Run production on Postgres.
Which n8n failures slip past an error workflow?
Anything that never runs, never ends or runs at the wrong time. An Error Trigger workflow fires only when an execution fails (n8n docs). A skipped schedule, a run with no timeout because EXECUTIONS_TIMEOUT defaults to -1 (n8n docs) and a Cron job on New York time all look healthy from inside n8n.
A workflow that never ran
Healthchecks.io gives each check a unique ping URL and alerts when a ping misses its period plus grace time (Healthchecks.io docs). End every scheduled production workflow with an HTTP Request node that pings one. This dead-man's switch catches the absence an Error Trigger cannot see. The heartbeat also flags trigger-node failures. Those do reach an Error Trigger, though without an execution ID, because the workflow never executed (n8n docs).
A run that never ends, a caller that gives up
EXECUTIONS_TIMEOUT stays at -1 until you set it. The per-workflow maximum defaults to 3600 seconds (n8n docs). Set an instance-wide timeout so a hung run ends on its own.
By default nginx closes a proxied connection after 60 seconds without data (nginx). A Webhook node that responds when the workflow finishes loses its caller on longer runs. Raise proxy_read_timeout, or respond immediately and do the work afterwards.
Schedules on New York time
GENERIC_TIMEZONE defaults to America/New_York (n8n docs). A Berlin team's 07:00 Schedule Trigger fires at 13:00 local time for most of the year (our arithmetic). n8n's Schedule Trigger troubleshooting page tells self-hosters to set GENERIC_TIMEZONE (n8n docs).
Credentials that fail a week late
This one is loud but late: an error workflow sees it only if one is attached and only a week after setup. Google's refresh tokens expire in 7 days when the OAuth consent screen is External and the app is still in Testing (Google). The n8n credential works at setup, then fails on an unattended run after day seven. Publish the app.
What changes when you run n8n in queue mode?
Executions move to worker processes fed through Redis. A dead worker stops being self-healing: since n8n 2.0, queue mode no longer retries stalled jobs (n8n changelog). The worker health endpoint is off by default (n8n docs). A CX33 holding main, Redis, Postgres and one worker costs 10.69 EUR a month (our arithmetic).
n8n's queue-mode page sets the conditions (n8n docs):
- Postgres, because n8n does not recommend queue mode on SQLite.
- The main instance's encryption key on every worker and webhook processor.
- No filesystem binary-data storage. S3, the documented alternative, is Business and Enterprise only (n8n docs).
- Worker concurrency of 5 or higher (the default is 10).
- Multi-main high availability only on Enterprise.
Set QUEUE_HEALTH_CHECK_ACTIVE=true on each worker and point an external check at it. After a crash, find the executions that never finished and re-run them by hand.
Do you need queue mode at all? n8n's benchmark reports up to 220 workflow executions per second on one main-mode instance with Postgres (n8n docs). That clears the Business plan's 40,000 monthly executions in about 3 minutes (our arithmetic). It is a best case on simple webhook workflows. Memory is the limit to watch: n8n "doesn't restrict the amount of data each node can fetch and process" (n8n docs). Add queue mode for a measured reason.
How often do you have to update self-hosted n8n?
At least once a month, by n8n's own guidance (n8n docs). The pace behind that advice is fast: n8n shipped 40 minor versions between 2.0.0 on 8 December 2025 and 2.40.0 on 15 September 2026, about one a week (our arithmetic on n8n's GitHub releases for 2.0.0 and 2.40.0).
Take a full backup before each one (n8n docs). Skipping updates is the expensive failure. GitHub's advisory GHSA-v98v-ff95-f3cp scored CVE-2025-68613 at CVSS 9.9 with fixes in 1.120.4 and 1.121.1 (GitHub). Three days after it was published, Censys observed 103,476 potentially vulnerable n8n instances (Censys, 22 December 2025).
Cyera rated Ni8mare (CVE-2026-21858) at CVSS 10.0. It was fixed in 1.121.0 on 18 November 2025 and disclosed on 7 January 2026 (Cyera), 50 days later by our count. An instance on a monthly update window had the fix before anyone knew to look for it.
Use the same window for n8n's built-in security audit, which reports an outdated instance (n8n docs).
The n8n production monitoring checklist
Twelve checks cover every failure above. Two endpoints ship switched off: /metrics needs N8N_METRICS=true and the worker health endpoint needs QUEUE_HEALTH_CHECK_ACTIVE=true (n8n docs). Every other row is a setting you choose once. A heartbeat on Healthchecks.io is free for 20 checks (Healthchecks.io).
| Check | Catches | Shipped default |
|---|---|---|
| Poll /healthz/readiness externally | Database disconnected or not migrated | /healthz alone ignores the database |
| Scrape /metrics | Trends for your own dashboards | Off; N8N_METRICS=true |
| Poll each worker's /healthz | A dead queue worker | Off; QUEUE_HEALTH_CHECK_ACTIVE=true |
| Error workflow everywhere | Failed executions | Set per workflow |
| Heartbeat on scheduled workflows | Runs that never happened | Not built in |
| GENERIC_TIMEZONE | Wrong-hour schedules | America/New_York |
| EXECUTIONS_TIMEOUT | Runs that never end | -1 |
| proxy_read_timeout above your longest synchronous webhook run | Webhook callers cut off | nginx 60s |
| Google OAuth apps published, not in Testing | Credentials dying on day seven | 7-day tokens |
| Save settings and pruning | Personal data, disk growth | all; 336 hours or 10,000 |
| N8N_ENCRYPTION_KEY off the box | Undecryptable restores | Auto-generated in ~/.n8n |
| Monthly update, backup first | Known vulnerabilities | Manual |
For the external poll, UptimeRobot's Solo plan checks every 60 seconds for 9 EUR a month billed annually (UptimeRobot). The workflow automation examples help decide which workflows deserve a heartbeat first.
When self-hosting n8n is the wrong call
When nobody owns the box. n8n Cloud Pro is 50 EUR a month billed annually for 10,000 executions (n8n), with nothing for you to patch. Against the 7.09 EUR CX23, self-hosting saves about 43 EUR a month (our arithmetic). If your hour costs more than 43 EUR, one hour of upgrade work a month cancels that saving.
At SME volume one well-configured box is not fragile. Most failures here are one-line fixes. A single CX23 with Postgres, the checklist above and a free heartbeat plan is a boring system, which is what production should be.
Self-host when someone already runs Linux servers, when workflow data must stay on infrastructure you control or when your volume outgrows Cloud Pro's 10,000 executions. Stay on Cloud when the monthly update would land on someone who has never opened a terminal.
The bottom line
Self-hosted n8n is a 7.09 EUR a month server bill plus a monthly maintenance habit. The habit decides whether the box is safe: Censys observed 103,476 potentially vulnerable instances three days after a CVSS 9.9 advisory (Censys). Every silent failure in this guide has a documented fix.
Set GENERIC_TIMEZONE, EXECUTIONS_TIMEOUT and N8N_ENCRYPTION_KEY on day one. NodeSparks builds owned automations in two lanes, SaaS replacement and ops automation. The Ops Automation Playbook ranks which workflows deserve a production box first.
Where a workflow is better written as code than dragged through a UI, that is AI agent development.
Frequently asked questions
How much does self-hosted n8n cost per month?
About 7.09 EUR a month excluding VAT for a single-box production install on a Hetzner CX23, by our arithmetic on Hetzner's published prices. The server is 5.49 EUR for new orders since 15 June 2026. Automated backups add a flat 20% (1.10 EUR) and a primary IPv4 address adds 0.50 EUR. A CX33 for queue mode comes to 10.69 EUR. The free Community edition adds no licence cost. Upgrade time is the line to price yourself, because n8n asks self-hosters to update at least once a month.
Which features does the free n8n Community edition leave out?
n8n's Community edition page lists log streaming, external secrets, Git version control, environments, multi-main mode, external storage for binary data, projects, custom variables, SSO (SAML and LDAP) and sharing as missing from the free edition. Free registration unlocks folders, debug in editor and custom execution data. Execution-data redaction is Enterprise-only. The Community edition runs under the Sustainable Use License, which permits use for your own internal business purposes. n8n's paid self-hosted Business plan costs 667 EUR a month billed annually and includes 40,000 workflow executions.
Does n8n store personal data in its execution logs?
Yes, by default. n8n saves the full data of every successful and every failed production execution, because both save settings default to 'all'. Pruning then deletes executions older than 336 hours (14 days) or beyond 10,000 in total, though annotated executions are never pruned. Hiding execution input and output through redaction is Enterprise-only on both n8n Cloud and self-hosted. On the Community edition the practical fix is to save less: turn success saving off for workflows that handle customer personal data and shorten EXECUTIONS_DATA_MAX_AGE.
What is n8n queue mode and do you need it?
Queue mode moves executions from the main n8n process to worker processes through Redis. n8n does not recommend running it on SQLite, every worker needs the main instance's encryption key and the worker health endpoint is off by default. Since n8n 2.0 a stalled job is not retried automatically. Most small teams can wait. n8n's own benchmark reports up to 220 workflow executions per second on a single main-mode instance with Postgres. That is a best case on simple webhook workflows, which is why memory is the limit to measure before adding workers.
Why did an n8n Google credential stop working after a week?
Google issues refresh tokens that expire after 7 days when an OAuth app has an External user type and is still in Testing status. A Google credential built on such an app in n8n works at setup, then fails on an unattended run after day seven. The fix is to publish the OAuth app in Google Cloud. Attach an error workflow as well, because the failed run only reaches a person if one is set.
How often should you update a self-hosted n8n instance?
At least once a month, which is n8n's own guidance, with a full backup before each update. n8n shipped 40 minor versions between 2.0.0 on 8 December 2025 and 2.40.0 on 15 September 2026, about one a week by our arithmetic. Skipping updates carries real exposure. On 22 December 2025 Censys observed 103,476 potentially vulnerable n8n instances exposed to CVE-2025-68613, a flaw GitHub's advisory scores at CVSS 9.9. Set N8N_ENCRYPTION_KEY yourself and store it off the box, because without that key a restored database cannot be decrypted.

