n8n AI Automation:
Self-Hosted Workflows for Beginners

What is n8n and why does self-hosting matter?

n8n (pronounced "n-eight-n", standing for "nodemation") is an open-source workflow automation platform. Like Make.com and Zapier, it lets you build automated workflows that connect apps and services. Unlike those platforms, n8n's source code is publicly available and you can run it on your own server — which means no per-operation fees, no monthly task limits, and full control over your data.

The n8n interface: similar to Make.com, more powerful

n8n uses a visual flow diagram interface similar to Make.com. Nodes (equivalent to Make.com modules) are connected by edges showing data flow. You build workflows by adding nodes, configuring their parameters, and connecting them. The primary difference is that n8n also supports JavaScript code nodes — you can insert a code block at any point in the workflow and write custom logic in JavaScript, which is executed server-side when the workflow runs. This makes n8n significantly more powerful than Make.com for complex logic without requiring you to build and deploy a separate application.

n8n's AI capabilities

n8n has dedicated AI nodes for the major providers: an OpenAI node (supporting Chat Completions, Assistants, and other OpenAI APIs), an Anthropic node (for Claude models), a Google Gemini node, and a Mistral node. It also has an "Agent" node — a higher-level abstraction that lets you build AI agents that can use tools and take multi-step actions. For most AI automation use cases, the OpenAI Chat Model node paired with a standard node chain is the right starting point. The Agent node becomes valuable when you want the AI to decide which tools to use and in what sequence.

Who should self-host n8n

n8n self-hosting is the right choice if: (1) you are comfortable with basic server management — SSH, running Docker commands, managing a VPS; (2) you run or plan to run automations at significant volume where Make.com's operation limits become a real constraint; (3) data privacy is important — your automations process sensitive client data, financial records, health information, or anything else you would rather not route through a third-party SaaS platform; or (4) you want the ability to write custom JavaScript logic within your workflows without paying for a professional plan on hosted platforms.

If the server setup sounds intimidating, start with Make.com and return to n8n when you are ready. The skills you build on Make.com translate directly to n8n — the concepts are the same; only the setup and the code-node capability differ.

Self-hosting n8n: complete VPS setup in 90 minutes

Here is the exact process I use to get n8n running on a fresh VPS. These instructions assume Ubuntu 22.04 LTS, which is the most widely used Linux distribution for this purpose and the one with the best-documented troubleshooting resources.

docker run -d \
  --name n8n \
  --restart always \
  -p 5678:5678 \
  -v ~/.n8n:/home/node/.n8n \
  -e N8N_HOST=YOUR_DOMAIN \
  -e N8N_PORT=5678 \
  -e N8N_PROTOCOL=https \
  -e WEBHOOK_URL=https://YOUR_DOMAIN \
  -e N8N_BASIC_AUTH_ACTIVE=true \
  -e N8N_BASIC_AUTH_USER=admin \
  -e N8N_BASIC_AUTH_PASSWORD=change_this_password \
  n8nio/n8n

Building your first n8n AI workflow

Now that n8n is running, here is how to build an email classification workflow — the same use case covered in the Make.com and Zapier guides, allowing direct comparison of the three approaches.

Connecting Gmail to n8n

In n8n, click "New Workflow." Click "Add first step." Search "Gmail" and select "Gmail Trigger." Click "Create New Credential." Select "OAuth2." n8n provides a callback URL — copy it. In the Google Cloud Console, create an OAuth 2.0 Client ID, add the n8n callback URL as an authorized redirect URI, and download the client credentials JSON. Back in n8n, paste the Client ID and Client Secret, then click "Connect" to complete the OAuth flow. Select the trigger event "Message Received" and configure it to watch your inbox.

The Gmail OAuth setup takes 15–20 minutes the first time and is the most complex part of the n8n setup. Once done, it is stored as a reusable credential that all your Gmail-connected workflows can use.

Adding the OpenAI node

After the Gmail trigger, click the "+" button to add a new node. Search "OpenAI" and select the OpenAI node. Click "Create New Credential," add your OpenAI API key, and save. In the node configuration: Resource = "Chat," Operation = "Message," Model = "gpt-4o-mini." In the Messages section, add a system message with your classification prompt and a user message mapping the Gmail trigger's email body and subject. Save the node.

Adding conditional routing with a Switch node

After the OpenAI node, add an "n8n" → "Switch" node. The Switch node evaluates the OpenAI response and routes to different branches based on its value. Configure conditions: if OpenAI output contains "BILLING" → Branch 1; if contains "SUPPORT" → Branch 2; if contains "SALES" → Branch 3; default → Branch 4. Connect different Gmail "Add Label" nodes or Slack notification nodes to each branch.

Adding the logging node

On each branch (or after the Switch via a Merge node), add a Google Sheets "Append or Update Row" node to log the classification. Map: timestamp = {{$now}}, email subject = Gmail trigger subject, category = OpenAI response output, email from = Gmail trigger from email.

Activating the workflow: Toggle the workflow to "Active" using the toggle in the top right. n8n will now trigger automatically when new emails arrive in Gmail. Unlike Make.com's 15-minute polling, Gmail triggers in n8n use the Gmail push notification API — emails are processed within seconds of arriving, not minutes.

n8n's AI nodes in depth: chains, agents, and memory

n8n has a more sophisticated set of AI-specific nodes than Make.com or Zapier, reflecting its focus on power users. Understanding these nodes helps you build more capable AI workflows.

The Basic LLM Chain node

The simplest AI workflow node — a prompt in, a response out. Connect it to any LLM node (OpenAI Chat Model, Anthropic Claude Model, etc.) and configure the system prompt. For straightforward classification and generation tasks, this is all you need and it is equivalent to what you do in Make.com with the OpenAI "Create a Completion" module.

The AI Agent node

A significantly more capable node that allows the AI to decide which tools to use and in what sequence to complete a task. You configure the agent with: a system prompt defining its role and goal, a set of tools it can use (other n8n nodes, HTTP requests, functions you define), and an LLM for reasoning. When the agent runs, it reasons about the task, calls tools as needed, evaluates the results, and continues until it has accomplished the goal or determined it cannot. This is the foundation for building genuinely agentic AI workflows — systems that plan and execute rather than just executing a fixed sequence.

Memory nodes

n8n provides several memory nodes that allow AI workflows to maintain context across runs: a Window Buffer Memory node (remembers the last N messages in a conversation), a Postgres Chat Memory node (stores conversation history in a PostgreSQL database for persistent long-term memory), and a Zep Memory node (integrates with the Zep memory management platform). These are valuable for building AI workflows that feel conversational — where context from previous interactions informs current responses.

The Vector Store nodes

n8n has native integration with several vector databases — Pinecone, Supabase pgvector, Qdrant, Weaviate — through dedicated nodes. This makes it possible to build proper RAG (Retrieval-Augmented Generation) pipelines entirely within n8n without writing code: a Vector Store Retriever node performs semantic search against your knowledge base, the retrieved content is passed to the AI node as context, and the AI generates a grounded response. This is a significant capability advantage over Make.com and Zapier, which require external tools or custom code to implement RAG.

5 practical n8n AI workflow examples

Workflow 1: Instant email classification and response (using Gmail push)

Unlike Make.com's 15-minute polling, n8n's Gmail trigger uses push notifications — new emails trigger the workflow within seconds. Node chain: Gmail Trigger → OpenAI (classify) → Switch (route by category) → [Gmail Label + Gmail Create Draft (for response categories)] + [Google Sheets Log]. This is the same workflow that takes 15+ minutes to process in Make.com's free tier, running in seconds in n8n.

You are an email triage specialist. Classify this email into exactly one category.

CATEGORIES: BILLING | SUPPORT | SALES | SCHEDULING | PERSONAL | SPAM | OTHER

Return a JSON object: {"category": "CATEGORY", "urgency": 1-5, "summary": "max 10 words"}
Return ONLY the JSON. No other text.

Workflow 2: AI agent for research and briefing

An AI Agent workflow that receives a company name via webhook, searches for recent news using the SerpAPI or Brave Search tool, retrieves the company's LinkedIn page via HTTP Request, synthesises the information into a structured briefing document, and saves it to Notion or Google Drive. This multi-step agent workflow — where the AI decides what information to gather and how to synthesise it — is difficult to build in Make.com without code but straightforward with n8n's Agent node.

Workflow 3: Customer FAQ responder with RAG

Node chain: Webhook trigger (receives customer question) → Embeddings (convert question to vector) → Pinecone Vector Store Retriever (find relevant FAQ chunks) → AI Node with retrieved context → HTTP Response (return the grounded answer). This is a proper RAG pipeline in n8n, keeping all data on your infrastructure and achieving significantly lower hallucination rates than prompting a model without knowledge base retrieval.

Workflow 4: Automated invoice processing with validation

Node chain: Gmail Trigger (watches for email with PDF attachments) → Extract Binary Data → HTTP Request to OpenAI Vision API (extract structured data from invoice image) → Code Node (validate extracted JSON against schema, flag low-confidence items) → Google Sheets (log valid items) → Gmail (send exception report for flagged items). The Code Node here handles the validation logic in clean JavaScript rather than forcing it through visual filter modules.

// n8n Code Node: Validate invoice JSON
const invoice = JSON.parse($input.first().json.choices[0].message.content);
const required = ['supplier_name','invoice_number','invoice_date','total_amount'];
const missing = required.filter(f => !invoice[f]);

return [{
  json: {
    ...invoice,
    is_valid: missing.length === 0 && invoice.confidence !== 'low',
    validation_errors: missing,
    needs_review: missing.length > 0 || invoice.confidence === 'low'
  }
}];

Workflow 5: Daily AI news digest with personalisation

Scheduled trigger (daily 7am) → HTTP Request to multiple RSS feeds (TechCrunch, The Verge, Wired, industry-specific sources) → Code Node (parse RSS XML, extract items from last 24 hours) → OpenAI (filter for relevance to your specific interests, summarise remaining items) → Gmail (send personalised digest). The Code Node for RSS parsing is a 10-line JavaScript function that handles what would take 5–6 Make.com modules to achieve.

Maintaining your n8n instance: the monthly checklist

Self-hosting requires some ongoing maintenance — not much, but enough to keep your instance secure and running reliably. Here is the monthly checklist that keeps a well-run n8n instance healthy with about 20 minutes of attention per month.

Monthly maintenance tasks

• Update n8n to the latest version. Run: docker pull n8nio/n8n && docker stop n8n && docker rm n8n, then re-run your original docker run command. n8n typically releases updates bi-weekly. Check n8n's GitHub releases page for breaking change notices before updating major versions.
• Check disk usage. Run: df -h on the server. n8n stores execution history which can grow. If disk is above 70%, either increase your VPS storage tier or configure n8n to prune old execution data (set N8N_EXECUTIONS_DATA_MAX_AGE=168 — keeps 7 days of history).
• Review error executions. In n8n's execution history, filter for "Error" status and review any recurring failures. A workflow consistently failing on the same input type needs prompt attention.
• Verify backup. Your n8n data is stored in ~/.n8n on the server. Confirm your backup system is running — either Hetzner's managed backups ($0.60/month) or a manual rsync to a backup location.
• Check API costs. Review your OpenAI and other API spending in their respective dashboards. Unexpected cost increases indicate either increased automation volume (good) or a runaway loop or prompt length inflation (needs investigation).

Frequently asked questions about n8n