§ Guide · n8n for teams
n8n turns scattered SaaS apps into one automated nervous system — with AI agents where it pays off.
n8n is more than a Zapier alternative. It's a visual and code-friendly workflow platform you can self-host, with native LLM and agent nodes. Below: what n8n is, when to choose it and how an Agentic Engineering team ships it to production.
01 — Definition
What is n8n?
n8n (pronounced 'n-eight-n') is a fair-code workflow automation platform that ties systems, APIs and AI models together without writing code for every flow. It uses a visual canvas of connected nodes: triggers, integrations, transformations and — since 2024 — fully fledged LLM and agent nodes.
The difference with Zapier or Make: n8n is self-hostable, code-friendly and has no per-task pricing. For teams that don't want data flowing through a US SaaS, or that need complex logic without vendor lock-in, n8n is often the natural foundation.
02 — Architecture
Core concepts on one page
Workflow
A directed graph of nodes that starts at a trigger (webhook, cron, event) and ends at an action or multiple outputs.
Nodes
400+ ready-made integrations (Slack, HubSpot, Postgres, Google, Microsoft) plus generic HTTP, Code and Function nodes for anything missing.
AI Agents & LangChain nodes
Native nodes for OpenAI, Anthropic, Gemini, embeddings, vector stores and tool-calling agents. Build a RAG or agent flow without a Python stack.
Credentials & Variables
Central management of API keys, OAuth tokens and environment variables — separated from the workflow definition so exports stay safe.
Executions & Queues
Per run you see every node's input/output, can re-run from any step, and self-host scales via a Redis queue + workers.
n8n versus the alternatives
| Dimension | n8n | Zapier | Make |
|---|---|---|---|
| Hosting | Cloud and self-host (Docker/K8s) | Cloud-only | Cloud-only |
| Pricing model | Per active workflow / self-host free | Per task per month | Per operation per month |
| Code freedom | JS/Python in nodes, custom nodes | Limited code steps | Limited filters/formulas |
| AI nodes | Native LangChain + agent nodes | Separate AI actions | OpenAI modules, no agents |
| Audience | Engineers + power-ops | Operators, marketing | Operators, prosumers |
| Data residency | Fully on your own network | US-bound | EU/US, no on-prem |
03 — Application
When to choose n8n (and when not)
Use n8n
- Operational automations with many integrations (CRM, ERP, ticketing, mail).
- AI flows that pull data from internal systems before an LLM call (RAG, enrichment).
- Sensitive or regulated data that can't flow through a US SaaS.
- Teams with DevOps capacity that want to decouple cost from volume.
Skip n8n
- A one-off, simple Slack-to-Sheets bridge — Zapier ships faster.
- High-frequency, low-latency event streams (>100/sec sustained) — use Kafka or a queue.
- Pure code projects with complex domain logic — write a Python service or edge function.
04 — Operations
Production, governance and cost
Self-host or cloud — a deliberate choice
Cloud is fast to launch; self-host (Docker/Kubernetes) gives full data residency and no task limits. We host by default in your own Azure/AWS/Hetzner with monitoring and backups.
Version control and environments
Workflows as JSON in git, with separate dev/staging/prod instances and credential mapping. No more 'who edited this on Friday' mysteries.
AI cost control
LLM nodes without budgets burn fast. We add token budgets, model routing (small model for classification, large for synthesis) and eval checks around every AI step.
On-prem AI via OpenClaw
For regulated clients we run n8n with a local inference layer via OpenClaw, so prompts and outputs never leave the network.
05 — Risk
Pitfalls we see in pilots
Workflow sprawl
Without naming conventions, tags and ownership n8n grows into 300 workflows nobody understands within a year. Start with a catalogue.
No error handling
By default a workflow fails silently. We add Error Trigger workflows, retries with backoff and alerting to Slack/Sentry on every production flow.
Credentials in workflow exports
Exported wrong, keys leak. Always use the credentials store and review JSON exports before they land in git or a ticket.
AI nodes without evals
An LLM step that looks great in a test run isn't an LLM step that performs consistently in production. Build eval suites and run them on every prompt change.
Our stance: n8n is the foundation, agents are the superstructure
n8n is our default orchestrator when multiple systems need to talk and a human or agent belongs in the loop. For pure agent-heavy work we combine with CrewAI or LangGraph; n8n stays the skeleton that handles triggers, data and delivery.
The difference between a fun pilot and a production automation always lives in the same places: hosting choice, observability, error handling, cost control and version control. That's Agentic Engineering.
n8n production-ready in your organisation?
We help teams set up or scale n8n: architecture, self-host on your own cloud, AI flows with budget control, monitoring and — if desired — an on-prem variant via OpenClaw.
Frequently asked questions
Is n8n really free?
+
The Community Edition is free to self-host without task limits. For cloud, SSO, RBAC and advanced features there are paid tiers. We usually self-host on the client's own infrastructure.
Can n8n fully replace Zapier and Make?
+
For most use cases yes, often with better cost structure and more control. For very simple, one-off bridges where non-technical users manage flows themselves, Zapier remains lower-friction.
How safe is n8n for confidential data?
+
Self-hosted, n8n runs fully on your network: data only leaves the environment for APIs you explicitly call. Combined with OpenClaw for local inference it fits strictly regulated data.
What's the difference between n8n and CrewAI?
+
n8n is a workflow orchestrator that can also run agents. CrewAI is a Python framework specifically for multi-agent collaboration. In production we combine them: n8n as skeleton (triggers, integrations, delivery), CrewAI for the actual agent work.