n8n Review 2026

How we evaluated this

We evaluated n8n as a workflow and AI-agent control plane, not as a chatbot builder. Scoring weights integration depth, workflow control (branch, loop, code, error, wait), deployment choice (cloud versus self-host), governance (Git, RBAC, SSO, audit, secrets), and meter honesty (executions versus tasks).

On August 19, 2026 we read n8n.io, n8n.io/pricing, n8n.io/pricing/startup-plan, n8n.io/vs/zapier, and the public comparison table (Starter / Pro / Enterprise columns). We did not obtain a live Enterprise quote. Your remaining tests: one production-shaped workflow at real trigger volume, a forced failure, and a written cloud-versus-self-host decision.

Pricing calculation to run

Count executions, then decide who hosts

Zapier-style tools charge per task. A webhook that writes ten spreadsheet rows and posts to Slack can be many tasks. n8n charges per completed workflow run. Official vs-Zapier copy: cloud tiers start at €20 per month for 2.5k workflow executions, unlimited users, unlimited workflows, every integration. Official pricing page: you pay when a workflow runs from start to finish, regardless of complexity. That is the original cost insight. Complexity is free. Frequency is not.

The second half of the calculation is hosting. Community Edition on your metal has no vendor execution fee. You pay VMs, storage, backups, monitoring, and the person who upgrades n8n. n8n Cloud trades that labor for a monthly execution quota and Microsoft Azure hosting in Germany West Central (Frankfurt), per official vs-Zapier security copy. Business, according to the Start-up Plan page, is currently self-hosted only. Enterprise can be hosted by n8n or self-hosted. If you pick Cloud, the meter is executions. If you pick Community, the meter is infra and people. If you pick paid self-host Business or Enterprise, you pay both a license that is still execution-scoped and the infra. Do not compare a €20 Cloud card to "free self-host" without adding labor.

• Forecast executions by trigger, not by node count. Cron every minute is 43,200 executions per month if it always completes. A form that fires 80 times a day is about 2,400, which already presses Starter's 2.5k. A 40-node workflow that runs twice a day is 60 executions. Node count does not save you. Frequency does.
• Separate editor runs from production. Unstable workflows during rollout can burn Cloud quota. Confirm what n8n counts as a billable production execution in your plan FAQ before you test in the paid workspace.
• Model retries as new executions when they are. Error workflows and automatic retries are listed on the pricing table. If a flaky API retries into a full rerun, your 2.5k or 40k number is not the happy-path count.
• Sub-workflows can count. Community discussion and vendor FAQs have treated called workflows as their own executions. Confirm the current rule on the live pricing FAQ. If each child run bills, fan-out designs get expensive on Cloud and stay cheap on Community hardware.
• Self-host is not "unlimited free Business." Community is free and capable. Paid self-host unlocks Git versioning, environments, SSO, queue mode, and support. The Start-up Plan is 333€ per month billed annually, 40k executions, 50 percent off Business, for teams under 20 employees and under €5M funding, for one year or longer if you still qualify. That math implies Business at 666€ per month billed annually for the same 40k. Confirm the live Business card. Monthly billing exists and will be higher.

A worked example using only official numbers. A support-triage workflow with 18 nodes that fires 3,000 times a month is 3,000 executions on n8n, over Starter's 2.5k, so you need Pro or a self-host path. The same flow on a task-priced tool could be 3,000 times N steps. If N is 18, that is 54,000 tasks. That is why n8n wins on paper for fat workflows. Flip the example: a one-node cron that hits a health-check URL every minute is 43,200 n8n executions, which is already above the 40k Business bundle. On Zapier that might be 43,200 cheap tasks. Execution pricing punishes chatty triggers and rewards dense graphs. Design the workflow to the meter, or the meter will design your architecture.

Who n8n fits best

Surfaces to verify

n8n is not a social inbox. The surfaces that matter are editor, API, CLI, cloud, self-host, and how workflows get invoked (webhook, cron, queue, MCP). Test the ones you will actually run.

• n8n Cloud editor hosted by n8n, with OAuth connectors and one-click upgrades.
• Self-host via Docker, npm, or your orchestrator. Community is free. Business and Enterprise unlock Git, environments, and governance.
• API and CLI to control instances. Extra API scope control sits on higher plans.
• Webhooks and queues as first-class triggers on all plans, including Cloud (vs-Zapier notes webhooks are core, not a premium add-on).
• Native app nodes plus HTTP Request for everything else. Confirm the exact node you need. A listed integration can still be thin.
• AI and MCP: call models from a workflow, or expose a workflow as an MCP tool to Claude or another client. That is orchestration, not a support-desk product. A tool gateway is a different job.
• Git and environments on paid self-host / Enterprise: dev, staging, prod, workflow diffs.

Workflow execution and guardrails

A production n8n workflow reads a system, optionally calls a model, branches on rules or confidence, waits for a human, writes to another system, and logs the run. The canvas is for structure. Code nodes are for logic the UI cannot express. Pair AI steps with deterministic checks before any write that touches money, access, or customers.

Configure error workflows, retries for transient failures, and human-in-the-loop nodes for risky actions. Restrict credentials with RBAC. On Enterprise, add audit logs and log streaming to your SIEM. Evaluations exist so you can score prompts instead of hoping. Treat n8n as a control layer for AI, not as an unsupervised agent.

• Lead capture and routing from forms or CRM webhooks.
• Support triage that classifies and opens a ticket, not a replacement helpdesk.
• Infra alerts that open incidents and page on-call.
• Document intake with RAG and a human verify step.
• Internal API orchestration across services you already run.

Official product demo

See the n8n canvas before you trust the execution claims

This official n8n quick-start shows the visual builder. Treat it as a tour. After watching, rebuild one of your real triggers, including a failure, and count the executions.

Pricing checked August 19, 2026

n8n pricing: Community free, Cloud from €20/mo, Business self-host, Enterprise custom

All paid plans include unlimited users, unlimited workflows, and every integration. The meter is monthly workflow executions. Cloud cards are JavaScript-rendered. Figures below are only those we could tie to official pages on August 19, 2026. We do not invent a Pro euro amount the static HTML did not print.

Source: n8n pricing, Start-up Plan, and n8n vs Zapier. Confirm monthly versus annual Cloud prices, Pro's euro amount, overage behavior, and AI Assistant credit rules before signing. Eligible startups under 20 employees and €5M funding can take 50% off Business for at least one year.

AI capability and agent execution

n8n treats AI as nodes inside a workflow. You can attach OpenAI, Anthropic, Gemini, Groq, local models via Ollama, vector stores, and evaluations. Multi-agent patterns exist (an orchestrator directing specialist agents). MCP support lets n8n call tools or expose workflows as tools to other clients. The AI Workflow Builder can draft a canvas from a description. None of that removes the need for structured outputs and human gates on writes.

The risk is drift. Without prompt versioning, eval cases, and an approval node, an agent step is hard to explain after it mis-routes a refund. Keep AI inside deterministic branches. If you only needed a support chatbot, buy a support product. If you only needed a model, buy Claude or ChatGPT and call them from a thin webhook.

Feature areas to verify

• Branch, merge, loop, wait, and error workflows on your data.
• JS and Python code nodes, including any library limits on Cloud versus self-host.
• HTTP, GraphQL, webhook, and cURL import for systems without a native node.
• Execution log search, debug-in-editor, and pin-data testing.
• Git versioning, environments, and workflow diffs if more than one person will edit.
• RBAC, shared projects, and instance admin.
• AI agent nodes, evaluations, and MCP trigger.
• Insights date range and granularity at your plan (Starter is days-scale, Enterprise is hourly to yearly).
• Queue mode and multi-main if you expect burst load. Some scaling rows are Enterprise or "coming soon."

Analytics and operating visibility

You get execution logs on all paid plans, plus Insights banners, tables, and time series with plan-gated filters. Starter Insights filtering is listed at 7-14 days. Enterprise lists 24 hours to 365 days and hourly granularity. The useful question is export: can failed runs stream to Datadog or your SIEM? Log streaming is an Enterprise row. If you cannot answer "which workflow failed at 02:14 and what payload did it see," you are not ready for production, regardless of node count.

Security, data handling, and compliance

n8n Cloud hosts in Microsoft Azure, Germany West Central, Frankfurt, per official vs-Zapier copy. Credentials are stored in an encrypted secrets store on paid plans. SSO SAML and LDAP, enforced 2FA, and external secret stores sit on Enterprise. Self-host shifts encryption-at-rest, backups, upgrades, and network isolation to you. Request the current SOC report and DPA rather than relying on older blog claims.

• Ask how long execution payloads, including error bodies, are retained on your plan.
• Map who can view executions that contain customer data.
• For self-host, write the backup, upgrade, and secret-rotation runbook before go-live.
• Verify audit-log coverage for workflow edits, credential access, and admin actions.
• If you need US data residency, Cloud-in-Frankfurt may be the wrong region. Confirm options in writing.

What questions should you ask before buying n8n?

• Can we build our top workflow on a trial with real triggers and real errors?
• What is the 12-month execution forecast including retries, cron, and seasonal peaks?
• Is Pro's live euro amount and monthly execution quota what we think it is?
• Cloud or self-host, and who pages when the instance or a workflow dies at 2 a.m.?
• Which connectors are native, which are community, and which are HTTP-only?
• How do Git, environments, and approval work for our team size?
• Do sub-workflow runs bill as separate executions today?
• Can we stream logs to our SIEM on this plan?
• What is the upgrade path from Community to Business if we start on our own VMs?

What red flags should you watch for with n8n?

• The demo never uses your trigger volume or your failure cases.
• The quote is a Cloud subscription with no execution forecast and no retry assumption.
• The team cannot name an owner for credentials, upgrades, or broken workflows.
• A critical system exists only as an unmaintained community node.
• AI write-actions have no evals and no human gate.
• Someone lists a tool gateway as an "n8n alternative" when the job is workflow automation. Those are different products.

What are the best alternatives to n8n?

If you want a simpler, task-priced SaaS builder, look at Zapier or Make. They are outside this review set and they price differently. Inside this site, pick by job, not by who has an Agent label. n8n is not listed against itself. A tool gateway is not an automation platform.

• YourGPTChoose YourGPT when the job is a managed support or sales agent across channels, not a general workflow engine you will operate.
• ClaudeChoose Claude when you need frontier reasoning and will attach it to systems yourself, without buying a visual automation suite.
• CursorChoose Cursor when the work is changing application code in an editor, not moving data between business apps.
• GitHub CopilotChoose Copilot when the need is pair programming and pull requests inside GitHub, not cross-app orchestration.
• ChatGPTChoose ChatGPT for ad-hoc analysis and drafts that a human will paste into existing tools.

Workflow test

What n8n needs to prove in a real automation

1. Trigger it.Fire from your real webhook, cron, form, or queue. Measure latency and whether each fire is one execution.
2. Transform it.Map fields, branch, and run a code node on real records. Look for silent nulls.
3. Guard it.Force an error and a low-confidence AI output. Confirm retry, error workflow, human approval, and alerting.
4. Observe it.Read logs and Insights. If you cannot debug a failed payload, do not put customer data on the canvas.

• Cursor review
• Claude review
• YourGPT review
• ChatGPT review
• How to choose an AI agent platform
• Tools directory
• Scorecard

Claim and source ledger

Should you choose n8n?

Choose n8n when you want to own automations end to end and you prefer paying for completed runs rather than steps. Start on Community if you can operate Docker. Start on Cloud Starter if you want n8n to host and you can stay near 2.5k executions. Move to Pro, paid self-host Business, or Enterprise when volume, Git, SSO, or an SLA shows up. Model chatty crons separately from fat graphs. Those two shapes have opposite economics on an execution meter.

Skip n8n if you have no technical owner, if you wanted Zapier simplicity, or if the job is a packaged support agent. In that last case use YourGPT. Run the four-step workflow test and a 12-month execution forecast before you sign anything annual.