n8n Workflow Automation: Build AI-Powered Workflows From Setup to Production

n8n workflow automation gives technical teams a visual, node-based editor for building everything from simple API integrations to full AI agent pipelines -- without locking you into a vendor's cloud. Created by Jan Oberhauser in Berlin, n8n (short for "nodemation") is a fair-code-licensed platform with 400+ integration nodes, optional JavaScript and Python code nodes, and native LangChain support for AI workflows. Over 205,000 GitHub stars. 100M+ Docker pulls. A $5.2 billion valuation after SAP's strategic investment. At this point, n8n is the default for teams that want to self-host their automation stack.

This guide covers setting up n8n, building your first n8n AI workflow automation, and figuring out where n8n shines versus where it starts to crack.

What Makes n8n Different From Zapier and Make

It comes down to control. Zapier charges per task (each individual action), Make charges per credit (one standard module execution equals one credit), and n8n charges per execution. A single workflow run counts as one unit no matter how many steps it contains. Got a 10-step workflow running 10,000 times per month? That pricing model can dramatically reduce costs compared to Zapier, which would bill for 100,000 individual tasks.

The bigger deal: self-hosted n8n has zero software-imposed execution limits. Your only constraint is your server hardware. That makes it dramatically cheaper at scale and gives you full data sovereignty. Your workflow data never leaves your infrastructure.

The tradeoff is ease of use. Zapier offers 9,000+ integrations with a pure no-code interface. n8n has roughly 400-500 native integrations and a steeper learning curve. Production-grade workflows typically require writing code in Code nodes, debugging API calls, and managing OAuth credentials by hand.

Setting Up n8n With Docker: Step-by-Step

The fastest path to a running n8n instance is the official one-line Docker installer. You'll need Docker Engine installed and a server with enough headroom (n8n recommends at least 2 GB of RAM for basic use, though 4 GB or more makes sense for production workloads).

Quick Start (Development)

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

This script verifies Docker is installed, creates an n8n/ folder, generates config files, and launches n8n with SQLite and the sandbox runner. It's idempotent, so re-running it is safe. The editor opens at http://localhost:5678.

To stop and start:

docker compose -f ./n8n/compose.yml down
docker compose -f ./n8n/compose.yml up -d

To upgrade:

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

If you want tighter control, try the manual Docker quickstart instead:

docker volume create n8n_data
docker run -it --rm --name n8n \
  -p 5678:5678 \
  -v n8n_data:/home/node/.n8n \
  docker.n8n.io/n8nio/n8n

The volume persists your workflows, credentials, and encryption key across container restarts.

Production Setup With PostgreSQL

For production, swap out the default SQLite for PostgreSQL and set an explicit encryption key. Create a .env file:

# .env
POSTGRES_USER=n8n
POSTGRES_PASSWORD=your-secure-password-here
POSTGRES_DB=n8n
N8N_ENCRYPTION_KEY=your-random-32-char-string
GENERIC_TIMEZONE=America/Los_Angeles
TZ=America/Los_Angeles

Critical warning: the N8N_ENCRYPTION_KEY encrypts all stored credentials at rest. Lose this key, and every credential in your instance becomes permanently unrecoverable. Set it before your first launch. Back it up somewhere secure.

For Docker secrets integration, n8n supports a _FILE suffix pattern. Append _FILE to environment variable names and it loads values from mounted secret files instead of plaintext env vars. Works with DB_POSTGRESDB_PASSWORD_FILE, N8N_ENCRYPTION_KEY_FILE, and other sensitive variables.

Never run docker compose down -v on a production instance. The -v flag deletes volumes, including your encryption key and database.

Building Your First n8n AI Workflow

n8n's AI capabilities are built on native LangChain integration. The key AI nodes:

• AI Agent: the orchestrator node that receives a prompt, reasons about which tools to use, and executes a multi-step plan
• AI Tool: exposes any n8n workflow or API call as a tool the agent can invoke
• AI Memory: adds conversation context (buffer memory, window memory, or vector store-backed memory) within a session
• Structured Output Parser: structures the agent's raw text output into JSON or other formats using a schema you define

Here's a minimal n8n AI agent workflow definition you can import directly. It creates a webhook-triggered agent that uses OpenAI and a calculator tool:

{
  "name": "Simple AI Agent",
  "nodes": [
    {
      "parameters": {
        "httpMethod": "POST",
        "path": "ai-agent"
      },
      "type": "n8n-nodes-base.webhook",
      "name": "Webhook Trigger",
      "position": [240, 300]
    },
    {
      "parameters": {
        "model": "gpt-4o",
        "options": {}
      },
      "type": "@n8n/n8n-nodes-langchain.lmChatOpenAi",
      "name": "OpenAI Chat Model",
      "position": [480, 460]
    },
    {
      "parameters": {},
      "type": "@n8n/n8n-nodes-langchain.agent",
      "name": "AI Agent",
      "position": [480, 300]
    },
    {
      "parameters": {},
      "type": "@n8n/n8n-nodes-langchain.toolCalculator",
      "name": "Calculator",
      "position": [640, 460]
    }
  ],
  "connections": {
    "Webhook Trigger": {
      "main": [
        [{ "node": "AI Agent", "type": "main", "index": 0 }]
      ]
    },
    "OpenAI Chat Model": {
      "ai_languageModel": [
        [{ "node": "AI Agent", "type": "ai_languageModel", "index": 0 }]
      ]
    },
    "Calculator": {
      "ai_tool": [
        [{ "node": "AI Agent", "type": "ai_tool", "index": 0 }]
      ]
    }
  }
}

Import this via Settings > Import from File in the n8n editor. Add your OpenAI credentials to the Chat Model node, activate the workflow, and POST a JSON body with a query field to http://localhost:5678/webhook/ai-agent.

You can extend this by adding more Tool nodes: HTTP Request tools for API calls, Code tools for custom logic, or the MCP Client Tool to call external MCP servers. For more on connecting n8n with the Model Context Protocol, check out our n8n MCP setup guide.

Real-World Use Cases for n8n AI Workflow Automation

n8n works best when you combine structured workflow logic with targeted AI decision-making.

Customer support triage. A webhook receives incoming tickets. An AI Agent node classifies urgency and intent. A Switch node routes to the right team queue, and an HTTP Request node updates the ticket system. Delivery Hero built a pattern like this for account recovery. One n8n workflow connecting Okta, Jira, and Google now handles roughly 800 monthly lockout requests that previously took 35 minutes each.

Lead enrichment. A CRM trigger fires when a new contact appears. Code nodes hit enrichment APIs (Clearbit, LinkedIn, company databases). An AI Agent node synthesizes the data into a lead score and summary, then the workflow writes the enriched record back to the CRM. StepStone, one of Europe's largest job platforms, runs 200+ production workflows on self-hosted n8n for similar data processing, cutting new integration time from two weeks to two hours.

Document processing. A file trigger watches a cloud storage folder. New PDFs get sent to an AI model for extraction (invoices, contracts, reports). Extracted fields are validated with Code node logic and pushed to a database or spreadsheet. Unbabel hit a 51% reduction in manual operational work by routing translation QA processes through n8n workflows.

n8n Self-Hosted vs. n8n Cloud vs. Gamut: Comparison

| Capability | n8n Self-Hosted | n8n Cloud | Gamut | |---|---|---|---| | Pricing model | Infrastructure cost only | Per execution (from EUR 20/mo) | Per agent | | AI agent support | LangChain nodes, within-session memory | Same as self-hosted | Persistent always-on agents with cross-session memory | | Integrations | 400+ nodes + HTTP node | 400+ nodes | 190+ native MCP integrations | | Hosting | You manage Docker/K8s | Managed (EU servers, Frankfurt) | Fully managed, zero-ops | | Scaling | Manual (queue mode requires Enterprise license) | Managed, with execution limits | Automatic | | Execution model | Trigger-based workflows | Trigger-based workflows | Always-on autonomous agents | | MCP support | Server trigger + client tool | Same as self-hosted | Native MCP across all integrations | | Setup effort | Hours to days (Docker, DB, TLS, monitoring) | Minutes | Minutes |

When n8n AI Workflows Hit Their Ceiling

n8n is great at structured, trigger-based automation. But as teams push into more sophisticated n8n AI agent territory, architectural constraints start showing up.

No persistent memory across executions by default. The built-in Simple Memory node stores context within a session but it's volatile. Data disappears when n8n restarts or when you save the workflow. Persistent memory means integrating external stores like PostgreSQL Chat Memory or Redis Chat Memory. If your agents need to learn from weeks of interactions, expect significant custom engineering.

Trigger-dependent execution. Every n8n workflow starts with a trigger: a webhook, a schedule, a database event. There's no concept of an always-on agent that autonomously monitors conditions and acts without being explicitly kicked off. Polling schedules can approximate this, but it's a fundamentally different architecture than persistent agent runtimes.

Scaling takes real DevOps work. Self-hosted n8n can struggle with large datasets. n8n doesn't restrict how much data each node can fetch and process, so workflows handling large volumes may cause memory errors. Horizontal scaling via queue mode requires an Enterprise license and PostgreSQL, plus Redis as a message broker.

Teams that have outgrown trigger-action chains and need agents that run persistently, maintain state across sessions, and orchestrate multiple AI models without infrastructure management might want to look at platforms like Gamut. It's a different model: always-on agents with native MCP integrations that can call n8n workflows (via MCP) while handling the autonomous reasoning and persistent memory that visual workflow tools weren't built for. For a broader look at automation platforms including n8n, see our best AI workflow automation platforms roundup.

FAQ

What is n8n workflow automation?

n8n is a fair-code workflow automation platform that lets you build automations visually by connecting nodes in a canvas editor. It supports 400+ integrations, JavaScript/Python code nodes, and native AI agent capabilities via LangChain. You can self-host it with Docker or use n8n Cloud.

Is n8n free?

The self-hosted Community Edition is free with no execution limits. It uses a Sustainable Use License (fair-code, not OSI open-source) that permits free use for internal business purposes but restricts commercial resale of the software. n8n Cloud starts at EUR 20/month for 2,500 executions (billed annually; EUR 24/month billed monthly).

Can n8n build AI agents?

Yes. n8n includes an AI Agent node with native LangChain integration, tool-calling capabilities, and within-session memory. The built-in memory is volatile by default, though. Persistent cross-session memory requires integrating external stores like PostgreSQL or Redis. Agents also can't run autonomously without a trigger event.

Does n8n support MCP?

n8n supports the Model Context Protocol with both an MCP Server Trigger (exposing workflows as tools for external AI assistants) and an MCP Client Tool node (calling external MCP servers from within workflows). It uses SSE or Streamable HTTP transport.

What happens if I lose my n8n encryption key?

All credentials stored in your n8n instance are encrypted with the N8NENCRYPTIONKEY. Losing this key makes every stored credential permanently unrecoverable. Set it explicitly before your first launch and store it in a secure vault.

Can I still install n8n with npm after version 3.0?

No. From n8n 3.0 (scheduled for October 2026), npm and npx installs are deprecated. Docker becomes the only supported installation method.