n8n
AI agents and workflows you can see and control
About n8n
n8n is workflow automation built for people who can write code but would rather not write all of it. You assemble a workflow on a canvas from several hundred integrations, drop into a JavaScript or Python step wherever the canvas runs out, and run the result on n8n Cloud or on your own infrastructure. What has changed about it recently is the agent layer: an AI Agent node, MCP server and client triggers, human approval for tool calls, and tracing so that each step of an agent reasoning run is visible on the same canvas as everything else.
That visibility is the actual argument for it. Most agent frameworks are code, which means debugging a misbehaving agent means reading a trace; n8n puts the run on a canvas where the branch that fired is the branch you can see. The commercial model is also unusual and worth understanding: n8n bills per full workflow execution rather than per step or per user, so a workflow with forty nodes costs the same as one with four, and every plan includes unlimited users and every integration. Be clear about the licence, though. n8n describes itself as fair-code, not open source. The Sustainable Use License restricts commercial use, files marked .ee. require a paid enterprise licence, and branches other than master are explicitly unlicensed. Source-available is the accurate description.
You build a workflow by chaining nodes: a trigger (a webhook, a schedule, a queue message, an MCP call), then integrations, conditionals, code steps and AI nodes. The canvas is the source of truth, and a code step is available at any point, so the common failure of low-code tools (everything is easy until the one thing you need is impossible) has an escape hatch.
Agents are nodes like anything else. The AI Agent node plans and calls tools, the MCP Server Trigger exposes your workflow to an external agent, and the MCP Client lets your workflow drive one. Human approval can be required before a tool call executes, which is the control most agent platforms leave to you to build.
Deployment is a genuine fork in the road rather than a detail. The Community Edition runs on your own hardware for nothing under the Sustainable Use License, while the cloud plans are metered by executions per month. Self-hosted and cloud share the same editor, and some features (LangSmith tracing, OpenTelemetry tracing) are available only in the self-hosted build.
- •Visual canvas with real code escape hatches - Several hundred integrations as nodes, plus JavaScript and Python code steps and arbitrary HTTP or GraphQL requests wherever a node does not exist.
- •AI agents as first-class nodes - An AI Agent node, AI steps, MCP Server Trigger, MCP Client and instance-level MCP, with each reasoning step traceable on the same canvas as the rest of the workflow.
- •Human approval for tool calls - A built-in gate requiring a person to approve before an agent executes a tool, rather than something you have to construct yourself.
- •Per-execution billing - Plans are metered by complete workflow executions with unlimited steps, so complexity is free and only the number of runs costs, which is a different trade from per-step or per-seat pricing.
- •Self-hosting as a first-class path - The Community Edition runs on your own infrastructure at no cost, and some tracing features exist only there.
- •Evaluation and tracing for AI steps - Metric-based evaluations on paid plans, with LangSmith and OpenTelemetry tracing available in the self-hosted build.
Technical teams who want automation they can inspect, and specifically teams who have outgrown Zapier or Make because they need a code step, a self-hosted deployment or a data-residency answer. It suits developers building internal agent workflows where the reasoning needs to be auditable, and it suits anyone whose automation volume makes per-step billing painful, since a complex workflow costs the same as a simple one here.
It is a poor fit if you want a genuinely open source licence, because the Sustainable Use License restricts commercial use and the enterprise files are not licensed at all. It is also awkward for a buyer outside the eurozone: the pricing page quotes euros only, with no USD figure anywhere, so what you actually pay depends on your card issuer. And the jump from Pro to Business is steep by any measure (60 EUR to 800 EUR a month for four times the executions), so the plan that fits a growing team may be considerably more expensive than the entry price suggests. Non-technical users who only need to connect two SaaS apps will find Zapier faster to reach an answer with.
Pricing
Priced in another currency
- Community Edition (self-hosted)$0/mo
- Starter24 EUR/mo
- Pro60 EUR/mo
- Business800 EUR/mo
- EnterpriseContact sales
From the vendor pricing page, 2026-10-16
