Building your first AI agent sounds complex, but with n8n and an LLM you can have a functional one in 30 minutes: it can make decisions, call tools, and return results. This guide takes you step‑by‑step, without unnecessary theory, to an agent you’ll actually use.
What an agent is and why n8n
An agent is not just a chatbot: it’s a flow that receives input, decides what to do with an LLM, executes actions (send an email, query an API, store data in a database) and returns a result. n8n is ideal for getting started because it’s visual, open source, and cheap to self‑host. You see every step of the flow, which makes it much easier to understand and debug what the AI decides.
What you need before starting
- A n8n instance (local with Docker or on a VPS).
- An OpenAI or Anthropic API key.
- The account of the service you want to automate (Gmail, Slack, a spreadsheet).
1. Installation in 5 minutes
The quickest way to try it out is to run n8n in Docker: a single command starts it locally. For production, a 5‑10 € per month VPS (Hetzner, DigitalOcean) with Docker Compose, a domain and an SSL certificate via Caddy or Traefik. That way your agent is available 24/7.
2. Your first AI workflow
Concrete example: a bot that classifies incoming emails and replies to the simple ones.
- Trigger: IMAP node connected to your Gmail.
- Classification: OpenAI/Anthropic node with a prompt that tags the email (question, spam, urgent…).
- If: branches the workflow according to the category.
- Response: node that generates a draft with the LLM and either sends it or leaves it in drafts.
3. Turn the workflow into a real agent
The jump from “workflow” to “agent” is made by n8n’s AI Agent node: instead of a fixed order of steps, you give it a goal and a set of tools, and the LLM decides which to use and in what order. For example: “answer the customer’s query” with access to a knowledge‑base search tool and another to create a ticket.
Examples you can set up today
1) Inbox assistant: classifies and replies to routine emails.
2) Daily summary: gathers updates from multiple sources and sends you a Slack summary each morning.
3) Lead capture: a form triggers an agent that enriches the contact and stores it in your CRM.
4) First‑level support: answers FAQs by consulting your documentation.
Common mistakes when starting
- Ambiguous classification prompts: define clear, closed categories.
- Not setting limits: make sure the agent doesn’t send anything without review until you trust it (use drafts).
- Forgetting the cost: each run calls the LLM; monitor spending on high volumes.
- Not handling errors: add failure branches so a weird email doesn’t break the whole workflow.
Our assessment
n8n is the best entry point to the world of agents for anyone who wants to understand what’s happening under the hood. In an afternoon you go from “agents sound like sci‑fi” to having one running on your own email. The key is to start with a small, real use case, not the ultimate agent.
Practical recommendation: set up the email classifier today in draft mode (no automatic sending). When you see it gets it right for a week, give it permission to reply on its own. That’s the safe path from prototype to production.
