n8n Workflow Automation: The Complete Guide for 2026

n8n workflow automation gives technical teams a visual canvas for connecting apps, APIs, and AI models, with full JavaScript and Python code access underneath. Jan Oberhauser started the project in Berlin in 2019. The name is short for "nodemation." Since then it's grown from a niche open-source tool into a platform with over 200,000 GitHub stars and a $5.2 billion valuation after a strategic investment from SAP in May 2026. This guide covers what you actually need to know if you're evaluating n8n for the first time or scaling existing workflows to production: Docker setup, AI workflow patterns, and the architectural limits that matter at scale.

What Is n8n and Why It Matters

n8n is a fair-code licensed workflow automation platform with native AI capabilities. You build automations by connecting nodes on a visual canvas. Each node represents an action: pulling data from an API, transforming a payload, sending a message, calling an LLM. When the visual editor can't express the logic you need, you drop into a Code node and write JavaScript or Python directly.

The core concepts are straightforward:

• Workflow: A collection of nodes that automates a process. Every workflow runs as a directed graph from trigger to output.
• Node: An individual component that fetches, sends, or processes data. n8n ships with 400+ built-in integrations.
• Trigger node: A special node that starts the workflow. Could be a webhook, a cron schedule, a new row in a database, or an incoming email.
• Credential: Stored authentication for external services. Credentials are encrypted and reusable across workflows.
• Expression: Inline JavaScript that dynamically populates node parameters using data from previous nodes.

n8n is available as a free self-hosted Community Edition, a managed Cloud service (starting at EUR 20/month billed annually), or paid Pro and Enterprise tiers. The self-hosted option is what makes it popular with engineering teams. You keep full control of your data and pay only for infrastructure, which typically runs $5-40/month on a VPS.

Setting Up n8n for Production

Most tutorials stop at npx n8n. Here's how to set up an instance you can actually rely on.

Docker Compose with PostgreSQL

The recommended production setup uses Docker Compose with PostgreSQL. A common recommendation for production is 4 GB RAM and 2 vCPUs, though n8n's official docs list 2 GB as the minimum.

The fastest path is n8n's one-line installer:

curl -fsSL https://get.n8n.io | sh

This creates an n8n/ directory with a compose.yml, auto-generated secrets, and a bundled search tool (SearXNG). It starts n8n at http://localhost:5678 with SQLite. Fine for testing. Can struggle under concurrent writes at higher volumes, though.

For production, use PostgreSQL. A minimal Docker Compose configuration looks like this:

services:
  postgres:
    image: postgres:18
    restart: always
    environment:
      POSTGRES_USER: ${POSTGRES_USER}
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD}
      POSTGRES_DB: ${POSTGRES_DB}
      PGDATA: /var/lib/postgresql/data
    volumes:
      - db-storage:/var/lib/postgresql/data
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -h localhost -U ${POSTGRES_USER} -d ${POSTGRES_DB}"]
      interval: 5s
      timeout: 5s
      retries: 10

  n8n:
    image: n8nio/n8n
    restart: always
    ports:
      - "5678:5678"
    environment:
      DB_TYPE: postgresdb
      DB_POSTGRESDB_HOST: postgres
      DB_POSTGRESDB_PORT: 5432
      DB_POSTGRESDB_USER: ${POSTGRES_USER}
      DB_POSTGRESDB_PASSWORD: ${POSTGRES_PASSWORD}
      DB_POSTGRESDB_DATABASE: ${POSTGRES_DB}
      N8N_ENFORCE_SETTINGS_FILE_PERMISSIONS: "true"
    volumes:
      - n8n_data:/home/node/.n8n
    depends_on:
      postgres:
        condition: service_healthy

volumes:
  db-storage:
  n8n_data:

One gotcha: the Postgres 18 Docker image changed its default data directory from /var/lib/postgresql/data to /var/lib/postgresql/18/docker. You must explicitly set PGDATA=/var/lib/postgresql/data or the volume mount misses existing data and the database starts empty.

Security Hardening

For any instance exposed to the internet, set these environment variables:

• N8N_SECURE_COOKIE=true -- HTTPS-only session cookies
• N8N_BLOCK_ENV_ACCESS_IN_NODE=true -- prevents Code nodes from reading server environment variables
• N8N_RESTRICT_FILE_ACCESS_TO -- semicolon-separated list of allowed directories (defaults to ~/.n8n-files in n8n 2.0+)
• MFA enforcement -- set N8N_SECURITY_POLICY_MANAGE_FROM_ENV=true and N8N_MFA_ENFORCED_ENABLED=true to force two-factor authentication for all users (available from n8n 2.18.0)

Use the _FILE suffix pattern for secrets (e.g., DB_POSTGRESDB_PASSWORD_FILE=/run/secrets/db_pass) to avoid passing credentials as plain environment variables. This integrates cleanly with Docker Secrets and Kubernetes secret mounts.

n8n AI Workflow Automation Patterns

In 2024, more than 80% of workflows built on n8n involved AI agents. That's driven by native LangChain integration and a dedicated AI Agent node. Here are the patterns that hold up in production.

Prompt Chaining

Connect multiple LLM calls in sequence, where each node's output feeds the next node's prompt. A typical chain: extract key entities from raw text, classify intent, then generate a structured response. Each step lives in its own node. You can swap models, adjust prompts, or add branching logic without touching the rest of the flow.

RAG with Vector Stores

n8n supports retrieval-augmented generation through vector store nodes. The pattern: ingest documents into a vector database (Pinecone, Qdrant, Supabase), then at query time retrieve relevant chunks and pass them as context to an LLM node. Works well for internal knowledge bases and support documentation.

Routing and Classification

Use an AI node to classify incoming data (support tickets by urgency, leads by intent, content by topic) then route to different workflow branches using n8n's Switch or If nodes. This is where n8n ai workflow automation really pays off. The classification is handled by an LLM, but the routing logic stays deterministic and auditable.

Human-in-the-Loop Approval

For any AI output that touches customers or production data, insert a Wait node that pauses execution until a human approves in Slack, email, or a custom form. This isn't optional for high-stakes automation. AI models hallucinate. Without guardrails, incorrect outputs in automated workflows can cause real damage: data corruption, malformed customer communications, worse.

Real-World n8n Workflow Examples

The value of n8n workflow automation gets concrete when you look at actual production deployments:

• Lead enrichment and routing: Trigger on a new HubSpot contact, enrich via Clearbit or Apollo, score with an LLM classification node, and route to the right sales rep in Slack. Runs in under two seconds per lead.
• Support ticket triage: Webhook receives a Zendesk ticket, an AI node classifies urgency and topic, the workflow assigns it to the correct team and fires a Slack alert for critical issues. Delivery Hero used n8n to automate 800 monthly account recovery requests, reducing total employee lockout time by 200 hours per month.
• Content pipeline: RSS trigger pulls new articles, an LLM summarizes them, another node drafts social posts, and a Wait node queues them for human approval before posting to LinkedIn and Twitter.
• Cross-department onboarding: HR system fires a webhook when a new hire is created. The workflow provisions accounts in Google Workspace, assigns Slack channels, creates Jira tickets for IT equipment, and notifies the hiring manager. All from a single trigger.
• Data sync and reconciliation: Scheduled trigger pulls invoices from Stripe, matches them against records in Airtable or a database, flags discrepancies, and sends a daily digest. StepStone cut per-source integration time from two weeks to two hours using n8n, with over 200 production workflows running on the platform.

n8n vs Zapier vs Make: When Each One Wins

The choice comes down to three factors:

Pricing at scale: n8n bills per workflow execution (the entire run counts as one), while Zapier bills per task (each step counts). A 10-step workflow running 10,000 times per month could cost 80-90% less on self-hosted n8n than on Zapier. Make falls in between with operation-based pricing.

AI capabilities: n8n has the deepest native AI integration of the three, with LangChain nodes, vector store connections, an n8n ai agent node, and MCP (Model Context Protocol) support. Zapier and Make offer AI features too, but they're bolted on rather than architecturally native.

Data sovereignty: If your compliance requirements mandate that automation data never leaves your infrastructure, n8n is the only option. It runs fully air-gapped if needed. Zapier and Make are cloud-only.

The trade-off: Zapier has 9,000+ integrations and the gentlest learning curve. n8n has roughly 400+ built-in integrations (plus thousands of community nodes) and expects you to be comfortable with JSON, APIs, and basic code. Pick accordingly.

Scaling n8n to Production

Default n8n runs everything in a single process with SQLite. That works until it doesn't. Here's the scaling path:

1. Migrate to PostgreSQL early. SQLite locks under concurrent writes and isn't supported in queue mode.
2. **Enable queue mode** with Redis. Set EXECUTIONS_MODE=queue to separate workflow triggering from execution, letting you add worker containers that process jobs in parallel.
3. Set execution concurrency limits. The environment variable N8N_CONCURRENCY_PRODUCTION_LIMIT controls how many production executions can run simultaneously. Configure it based on your instance's RAM to avoid out-of-memory crashes.
4. Prune execution history. Old execution data accumulates fast. Set EXECUTIONS_DATA_MAX_AGE (default: 336 hours) and EXECUTIONS_DATA_PRUNE_MAX_COUNT (default: 10,000) to keep the database lean.
5. Monitor externally. n8n's built-in observability is minimal. Export metrics to Prometheus or Datadog and set alerts on execution failures, queue depth, and memory usage.

When Workflow Automation Hits Its Ceiling

n8n is excellent at what it does: connecting systems through event-driven, node-to-node flows. But there's a class of problems where triggered workflows structurally can't help.

n8n workflows are stateless between executions. Each run starts fresh with no memory of previous runs unless you explicitly persist state to an external database. The n8n ai workflow pattern can call an LLM within a single execution, but it can't maintain a reasoning thread across multiple interactions, re-plan when conditions change mid-task, or autonomously decide which tools to use based on evolving context.

Think about investigating a production incident. The agent needs to check logs, form a hypothesis, query a database to test it, revise based on results, check a third system, write up findings. That requires persistent memory, dynamic tool selection, and multi-step reasoning that adapts at each step. A workflow can handle a fixed, predetermined path through those steps. It can't handle the case where step three's results invalidate the plan and require a completely different approach.

This isn't a knock on n8n. It's a fundamental distinction between workflow orchestration and autonomous agents. Different tools for different problems.

For teams that hit this boundary, platforms like Gamut provide persistent AI agents that maintain state across interactions, reason through multi-step tasks, use tools dynamically, and adapt their approach based on intermediate results. The practical pattern is to use n8n for deterministic, event-driven automation (it's genuinely great at this) and hand off complex reasoning tasks to autonomous agents that can operate independently.

FAQ

Is n8n free or open source?

n8n is free to self-host under the Sustainable Use License. That's source-available but not traditional open source (not MIT or Apache). The Community Edition includes unlimited workflows and executions. Cloud hosting starts at EUR 20/month billed annually. Enterprise features like SSO, SAML, and LDAP require a paid license.

Can n8n build AI agents?

n8n has an AI Agent node with LangChain integration, memory nodes, and support for MCP. These work well for single-execution AI tasks like classification, summarization, and RAG queries. But n8n agents are fundamentally workflow orchestrators. They follow predetermined node paths and lack persistent memory between executions. For tasks requiring ongoing reasoning and dynamic replanning, purpose-built agent platforms are a better fit.

How does n8n pricing compare to Zapier at scale?

n8n uses execution-based pricing (one workflow run = one execution regardless of how many nodes it has), while Zapier charges per task (each node step counts). At high volume, self-hosted n8n can cost 80-90% less. n8n Cloud narrows the gap but remains cheaper than Zapier for multi-step workflows.

What are the system requirements for self-hosting n8n?

A common production recommendation is 4 GB RAM and 2 vCPUs with Docker and PostgreSQL, though n8n lists 2 GB as the documented minimum. A basic VPS from DigitalOcean or Hetzner at $10-20/month handles most workloads. Official deployment guides cover AWS, Azure, GCP, and several other providers.

What happens to npm installs after n8n 3.0?

n8n 3.0, scheduled for October 2026, drops npm distribution entirely. Docker becomes the only supported installation method. n8n version 2 will remain supported for a year after the release of version 3. If you're currently running n8n via npm, plan your migration now.

Does n8n support MCP (Model Context Protocol)?

Yes. n8n has built-in MCP Server Trigger and MCP Client Tool nodes. The MCP Server Trigger lets n8n act as an MCP server, making its workflows available to AI assistants like Claude Desktop. The MCP Client Tool lets n8n agents call external MCP-enabled tools from within workflows. This makes n8n a useful bridge between AI agents and the 400+ services n8n integrates with.