What n8n actually is in 2026
n8n in 2026 is a fair-code workflow platform built for technical teams. You wire nodes on a canvas, drop into code when a step needs it, and connect hundreds of apps. It runs self-hosted for free or in the cloud, metered by workflow execution. It has also grown AI features: agent nodes and model steps you place in the graph like any other node.
So most people searching for an n8n alternative are not missing a feature. They are tired of being the person who does the wiring. n8n hands you total control, and total control means you design every branch, catch every error, and fix every broken node yourself. For internal plumbing that trade is often right. For growth work, it is worth questioning.
Workflow or agent: where the judgment lives
A workflow is a fixed graph. The trigger fires, data moves down branches you predicted, actions you wired execute. It never weighs anything. If a thread needs reading before you reply, or a lead needs scoring before you write, that judgment has to be a node you built in advance — and every case you did not predict becomes a silent failure or a broken branch.
An agent carries the judgment inside the run. You state the job; it reads the thread, checks memory for what worked before, drafts the reply, and stops at the approval gate with the price shown. Nothing sends, posts, or publishes without your yes. n8n's AI nodes put a model inside the graph — but the graph is still yours to design, test, and repair.
A migration path for growth workflows
You do not migrate n8n workflows so much as retire them. The candidates are the growth ones: monitoring sweeps, reply and outreach batches, keyword reports, page production, competitor checks. Internal plumbing — syncs between your own systems, webhook glue between databases — stays in n8n. It was never the target, and forcing it over would be the wrong move.
Each fire shows its price before it runs. A monitoring sweep is 1 credit; a reply or outreach batch of five drafts is 1 credit; a written and formatted page is 3. If a run fails during the switch, the credit comes back automatically, so a rough first week costs you nothing but the reruns.
- Write each growth workflow as one sentence in chat — conversation is free, so scoping the whole migration costs nothing
- Let the agent ask for accounts as runs need them — one OAuth popup per connection, no passwords handed over
- Run old and new in parallel for one cycle and compare outputs before trusting the run
- Retire the n8n workflow only when the run's results are on your board and holding
The honest cost shape: self-host hours vs credits
Self-hosted n8n is free software with a hidden invoice. You pay for the server, and someone technical pays in hours: version upgrades, nodes that break after an API change, error queues nobody watched over the weekend. Cloud n8n removes the server and meters flat tiers by execution — predictable, but the wiring and the repair time stay yours either way.
Moatt's shape is credits per run, with the price shown at the approval gate before anything fires. Conversation is free. Unused credits roll over up to twice your monthly amount, failed runs refund themselves, and overage bills at 1.25x capped at 3x, with alerts at 50, 80 and 100 percent. You pay for finished jobs, not executions.
Be honest about volume: if your automations are thousands of pure data moves a day, execution tiers or a self-hosted box will beat credits on price per fire. Credits win when the run replaces work — reading, scoring, writing, deciding — not when it merely moves fields between apps.
When n8n stays the right tool
Keep n8n when the job is integration, not judgment. Arbitrary connections between internal systems, data that must stay on your own hardware, high-volume moves where per-execution cost matters, or a technical team that genuinely likes owning the graph — n8n is excellent at all of it, and its community edition costs nothing to run.
The common end state is both: n8n keeps the plumbing, the agent takes the growth jobs where a decision sits inside the run. If nothing in your automation stack reads, scores, writes, or waits for approval, you may not need an agent yet — and it is fine to say so.