The question shows up every few weeks now, usually phrased as a shrug: is n8n dead? Someone spends a weekend wiring nodes together, then watches a coding assistant produce a working Python script for the same job in four minutes. The conclusion lands hard. Why drag boxes around a canvas when the model will just write the thing?

Short answer: n8n is not dead, and the evidence is boring and checkable. The repository ships releases most working days, it raised a large round in late 2025, and the connector library keeps growing. What is dying is the middle of the market visual automation used to own by default. Trivial automations are being absorbed into the apps themselves. Complex ones are increasingly written as code, because writing code got cheap. The band in between is narrower than it was in 2022, but it is real, and it is where a workflow tool still earns its keep.

So the useful question is not whether n8n survives. It is whether a visual workflow tool is the right container for this specific automation. That has an answer, and seven things decide it, none of which is how much you enjoy the canvas.

What n8n actually is, and the licence detail almost everyone gets wrong

n8n is a workflow automation tool you can self-host, where each step is a node on a canvas and any node can be swapped for a JavaScript or Python block when the visual abstraction stops helping. Its integrations page currently lists 2,174 integrations.

Licensing is where the community discussion goes wrong most often. n8n is not open source in the OSI sense. It ships under the Sustainable Use License, which n8n created in 2022 and describes as “fair-code”: source openly available, anyone can extend it, commercial use restricted by the authors. The licence grants rights to use, copy, distribute and make derivative works “for your own internal business purposes or for non-commercial or personal use,” and permits distribution only “free of charge for non-commercial purposes.”

n8n’s docs draw the line like this. Allowed: syncing your company’s data internally, building your own nodes, consulting around it, embedding n8n in your application using your credentials. Not allowed: white-labelling n8n as a commercial product, hosting it and charging customers for access, or collecting your users’ credentials to power a third-party app feature. Separately, any file with .ee. in its name or .ee in its path sits outside the Sustainable Use License and needs a valid n8n Enterprise License, defined in LICENSE_EE.md.

This matters more than licence pedantry usually does. If you are an agency planning to resell hosted n8n to clients, read the licence before you build the business model. And if you are self-hosting internally, know what you are not getting free: n8n’s docs list SSO (SAML and LDAP), log streaming, environments, Git-based source control, external secrets, projects, workflow and credential sharing, custom variables, multi-main mode and external binary storage as paid. Queue mode and standard logging are free. A no-cost “registered” community edition, unlocked by entering your email under Settings > Usage and plan, adds folders, debug-in-editor and custom execution data.

Git-based source control being enterprise-only is the one that bites engineering teams hardest. Hold that thought.

Is the project healthy? What the record shows

Checkable signals, September 2026:

• Release cadence. The GitHub releases page shows n8n 2.38.7 published on 11 September 2026, 2.39.x pre-releases shipping the same week, and the 1.x line still getting patches (1.123.79 on 10 September 2026). A maintained stable branch, an active next branch, and backports.
• Repository. Roughly 204,000 GitHub stars. Stars measure attention rather than health, but this is not the profile of an abandoned project.
• Funding. n8n announced a $180m Series C on 9 October 2025, led by Accel at a $2.5bn post-money valuation, taking total funding to $240m. Named investors include Meritech, Redpoint, NVIDIA’s NVentures and T.Capital, with follow-on from Sequoia, Felicis and Highland Europe.

The same announcement claims “6x user growth, 10x revenue growth” for the year. That is the company describing itself with no methodology attached, so treat it as marketing. The release log and the funding are verifiable; the multiples are not. None of this proves n8n is right for you. It does mean “dead” is the wrong word, and that platform risk here is low relative to most smaller competitors.

What LLMs really took away

Be precise about the threat, because the vague version produces bad decisions. Coding assistants did not make integration work disappear. They made the first draft of it nearly free. Different things. A model will write you a Stripe-to-Postgres sync in one pass. It will not notice, eleven weeks later, that Stripe deprecated a field and your nightly job has been silently writing nulls. That gap between working code and code that keeps working is where the value of a workflow platform lives, and LLMs barely touched it.

What they did kill is the bottom rung. “When a form is submitted, post to Slack” is now built into the form tool. Every serious SaaS product ships native integrations with the four or five apps its customers actually use, plus a webhook. If your automation is one hop between two popular products, you probably need no automation platform at all, and paying for one is a tax on not having checked.

At the top, the calculus flipped. A five-branch pipeline with retries, idempotency keys and backfill used to be weeks of engineering, which made a visual tool attractive even when it fit badly. A competent engineer with an assistant now produces that in a day or two, so the visual tool has to win on merit rather than on effort saved.

Four options, and the criteria that decide

Your realistic choices: a native integration in the app itself, a hosted no-code tool like Zapier or Make, a self-hosted workflow tool like n8n, or a script you write and deploy. What separates them:

1. How many integrations, and who maintains them. Zapier advertises 9,000+ apps, Make 3,000+, n8n 2,174. If your automation touches six niche SaaS products, connector coverage dominates everything else and the hosted tools win. If it touches your own database and one well-documented API, connector count is irrelevant.

2. How you get billed, which is not a detail. The three meter fundamentally differently, and at volume the gap is enormous. Zapier counts a task per successful action (polling is free). Make charges credits per action, most costing one credit, routers and error handlers free. n8n counts an execution as one run of the whole workflow, and its pricing page is explicit: “It doesn’t matter how many steps are in the workflow or how much data it processes.”

Work that through. A workflow with six acting steps, run 20,000 times a month, is roughly 100,000 tasks on Zapier, roughly 100,000 credits on Make, and 20,000 executions on n8n. Published n8n Cloud tiers at the time of writing: Starter €20/mo (2.5K executions), Pro €50/mo (10K), Business €667/mo (40K), billed annually, Enterprise on request. Zapier’s free plan is 100 tasks a month, Professional starts at $19.99/mo and Team at $69/mo, with the real price set by your task tier. Make lists Free, Core $12/mo, Pro $21/mo and Teams $38/mo, credits on top. Self-hosted n8n has no per-execution charge at all; you pay for the server. For high-fan-out work this is the biggest cost lever you have, and it is settled before you build anything.

3. Error handling and retries. Ask what happens on the 3am failure. A visual tool gives you retry policy, error branches and a failed-run record without writing any of it. A script gives you whatever you bothered to build, usually a try/except and a log line nobody reads.

4. Observability and replay. The honest killer feature. Opening a run from last Tuesday, seeing the exact payload at step four, fixing the mapping and re-running that one item saves real engineering time. Reproducing it in code means structured logging, payload capture, a replay path and somewhere to store it all.

5. Data residency and control. If the data cannot leave your infrastructure, self-hosting is not a preference, it is the requirement. This is the clearest case where n8n beats Zapier and Make outright.

6. Who needs to be able to change it. If the person who understands the rule sits in ops or finance and the rule changes monthly, a canvas they can open is worth real money. If only engineers will ever touch it, the canvas is overhead.

7. Whether it needs software engineering practice. Code review, staging, automated tests, rollback? If yes, and you are on community-edition n8n where Git source control is paid, you are about to fight your tool for a year.

What visual tools do better than code, and what they do much worse

Genuinely better:

• Credential management. OAuth refresh loops, token rotation and encrypted storage, handled once and reused everywhere. Tedious, security-sensitive code you now do not write.
• Connector maintenance as somebody else’s job. When an API version changes, someone else updates the node. You upgrade.
• Run history and replay, as above.
• Non-engineer editability. Changing a threshold from 500 to 750 should not require a deploy, and on a canvas it does not.

Genuinely worse:

• Version control. Workflows are JSON blobs, diffing them is miserable, and on n8n the Git integration is an enterprise feature. Most teams end up with a folder of exports and a naming convention, which is not version control.
• Testing. There is no natural unit test for a node graph. You test by running it, which means testing in production more often than you would admit.
• Code review. You cannot meaningfully review a canvas in a pull request.
• Complex branching. Past roughly a dozen nodes with nested conditionals, the picture stops clarifying and starts hiding things. When you are squinting at a graph to trace a condition, the abstraction has already failed.
• Anything needing engineering discipline: shared libraries, typed contracts, dependency management, staged rollout.

A rough rule: under a day of engineering plus stable business logic means code. Five third-party APIs plus logic that changes monthly means the workflow tool. The awkward cases are both at once.

Where AI agents fit, and where they should not

Both n8n and the hosted platforms now ship agent nodes, and the temptation is to hand an LLM some tools and a goal and let it work out the steps. For open-ended work with a human reviewing output, that can be excellent. For a process running unattended a thousand times a day it is usually a mistake, and the reason is arithmetic. Failure compounds: a step that works 97% of the time, chained eight deep, gives a run that works about 78% of the time. The τ-bench paper makes the sharper version of the point by measuring consistency rather than best-case success: it introduced a pass^k metric across repeated trials and found agents “quite inconsistent (pass^8 <25% in retail),” with the leading function-calling models of the time succeeding on under half of tasks in a single attempt. Those numbers are from 2024 and models have improved a great deal since, so do not quote them as current capability. The structure survives: an agent that sometimes succeeds is not a process that runs.

The production pattern is narrower. Use the LLM for the step that needs judgement, usually classification, extraction or drafting, and keep everything around it deterministic. Categorise the inbound email with a model; route it with an if-node. Extract invoice fields with a model; validate against a schema and fail loudly on mismatch. Let the model draft the reply; require a human to send it above a threshold. The workflow tool’s job in the agent era is to be the deterministic scaffolding an LLM sits inside, a more defensible role than code substitute ever was.

Self-hosting: the bill nobody puts in the spreadsheet

“Free and self-hosted” compares badly to €50/mo once you count honestly. You are taking on a Node application, a Postgres database, and Redis if you run queue mode for concurrency. That means backups you have tested by actually restoring them, TLS certificates, an upgrade path across a project shipping releases several times a week, and monitoring for the automation platform itself, because a workflow engine that is silently down fails in the quietest possible way.

Budget a modest VPS, managed Postgres, a few hours of someone’s attention monthly in the steady state, and a bad afternoon when a major version lands. If your automations are worth less than that, use the cloud tier and stop. If they carry regulated data, or enough volume that per-execution pricing hurts, self-hosting pays for itself fast. The mistake is self-hosting to save €50 a month while quietly spending €400 of engineering time.

A checklist for one specific automation

Five minutes, before you build anything. It settles most arguments.

• Does the source app already do this natively? Check first. Half of all automation projects are a setting somebody did not find.
• How many third-party systems does it touch? One or two with good SDKs, lean toward code. Four or more, lean toward a workflow tool.
• How often does the logic change, and who changes it? Monthly, by a non-engineer, means visual. Rarely, by engineers, means code.
• How many runs a month, times acting steps per run? That product decides whether per-action billing is affordable or absurd.
• What happens if it fails silently for a week? If the answer is bad, you need run history and alerting from day one, and that is easier to buy than build.
• Can the data legally sit on a third-party vendor’s infrastructure? If not, the shortlist is self-hosted, and the licence terms above apply.
• Does any step need judgement rather than rules? That step gets an LLM and a validation gate. The rest stays deterministic.
• Could you live with no tests and no code review here? If not, and you are on community-edition n8n, plan around the Git limitation before you have forty workflows.

The honest summary: visual workflow automation lost its monopoly at the easy end and its excuse at the hard end. It kept the middle, where you want integrations you do not maintain, a record of every run, and a canvas the ops lead can open without filing a ticket. Smaller than the 2022 hype implied. Not dead.

Common follow-ups

If n8n is not open source, can I get locked in? Less than with Zapier or Make, more than with a genuinely open-source tool. You can run it yourself indefinitely and workflows export as JSON, but the enterprise-gated features (SSO, Git source control, environments, external secrets) are exactly the ones a growing team wants. Decide upfront whether you will pay for them or design around them.

Should we migrate off Zapier to self-hosted n8n to save money? Only if task counts are high and workflows are step-heavy, because that is where per-execution billing wins. Count your actual monthly tasks first. Below a few thousand, migration costs more in engineering time than five years of subscription saves.

Can I just have an AI write and maintain my n8n workflows? It will write them. Maintenance means noticing breakage, which requires monitoring and someone accountable. Generation was never the bottleneck.

If you are weighing this for a process that actually matters, AB7 Solutions does automation architecture of exactly this shape: auditing what you already run, deciding per workflow whether it belongs in n8n, a hosted platform or a small service, and building deterministic guardrails around the steps where an LLM genuinely helps. We also take the unglamorous half, which is self-hosting, monitoring and keeping connectors alive after launch. Call +1 321 341 7733, or email ab@ab7solutions.com or director@ab7solutions.com, and there is more on our automation and agentic AI work at ab7solutions.com.

Sources: n8n Sustainable Use License (LICENSE.md); n8n Docs: community licence and fair-code; n8n Docs: choose how to use n8n; n8n GitHub releases; n8n blog: Series C announcement, 9 Oct 2025; n8n pricing; n8n integrations; Zapier pricing; Make pricing; Yao et al., τ-bench: A Benchmark for Tool-Agent-User Interaction (arXiv:2406.12045).