What an AI agent is in n8n.
A classic automation follows a path written in advance: if this, then that. An AI agent receives a goal and chooses the actions to reach it. Faced with a customer email, it works out whether it is a complaint, an order or a question, looks up what it needs, then acts: it replies, creates a task or hands over to the right person.
In n8n, the agent takes the form of a central node, AI Agent, to which you connect a language model, tools and memory. n8n builds on LangChain concepts for this, and you do not need to know them to get started.
The difference with a chatbot is covered in AI Agent vs Chatbot. The short version: a chatbot answers, an agent acts.
The four building blocks of an agent.
| Block | Its role | Examples in n8n |
|---|---|---|
| Trigger | What starts the agent | A message in n8n's chat, an incoming email, a form, a schedule |
| Language model | The reasoning | OpenAI, Anthropic, Google and other providers |
| Tools | What the agent can do | Read or write in the CRM, send an email, call an API, start another workflow |
| Memory | What it remembers | A conversation's history or the context of a case |
Tools are the block that matters most. An agent is only as useful as what it is allowed to do in your systems. That is also where security and guardrails come into play.
The five-step method.
An agent that fails in production is almost always an agent launched before it was scoped. Here is the order we follow:
Scope one single task
An agent that does one thing well beats an agent that does everything halfway. Write in one sentence what it must achieve, and what it must never do.
Pick the trigger
Email, form, message, schedule: start from where the request already arrives today, so nothing changes in the team's habits.
Connect the model and write the instructions
Choose the model for the task and write the instructions: role, tone, decision rules, the cases where it must stop.
Give it tools, and only those
Each tool is a right to act. Start read-only, then open write access tool by tool, with human approval on sensitive actions.
Test on real cases, then monitor
n8n offers evaluations to measure how reliable an AI workflow is. Test on real examples before going live, then review errors every week.
The guardrails that make the agent reliable.
The most useful n8n feature for putting an agent into production is human approval for tools. It works like this:
- The agent decides it needs a tool that requires approval, for example sending an email to a customer.
- The workflow pauses and sends an approval request to Slack, Microsoft Teams, Telegram, Discord or n8n's chat.
- The reviewer sees which tool the agent wants to use, and with which parameters.
- They approve: the action runs. They deny: nothing is sent, and the agent is told.
You can apply it to all tools or only some. Good practice: human approval at the start on anything that writes, sends or deletes, then loosen it over the weeks as trust builds.
Add three simple rules: a clearly bounded scope of action, a complete trace of what the agent did, and an alert as soon as a run fails.
Three concrete agent examples.
Qualify inbound requests. A prospect fills in a form. The agent enriches the company from authorized sources, assesses the request against your criteria, creates the CRM record with a summary and alerts the right sales rep. The rep calls back knowing who they are talking to.
Prepare every prospecting message. You give the name of a prospect. The agent reads their profile and recent activity, then drafts a tailored first message that talks about what they actually do rather than a generic template. The sales rep reviews, adjusts and sends. It is an agent we use ourselves, every day, on our own targets.
Process invoices arriving by email. Among all incoming messages, the agent recognizes those containing an invoice, downloads it and files it in the right month. Accounting gets a complete folder, with nobody sorting attachments. It is our own back office, and it runs continuously.
n8n, agents and MCP.
MCP (Model Context Protocol) is the standard that lets AI tools connect to other applications. n8n supports it both ways:
- An n8n agent can call external MCP servers with the MCP Client Tool node, and use the tools they expose.
- An n8n workflow can become an MCP server with the MCP Server Trigger node. You expose exactly the tools you designed, nothing more.
- n8n has its own MCP server. Since version 2.13, you can create and edit workflows by describing them to Claude, ChatGPT or a coding agent.
To understand what MCP changes for a business, see MCP (Model Context Protocol). For a full overview of the tool, see the n8n guide.
AI agent with n8n: what people ask us.
Can you build an AI agent with n8n without coding?
Mostly, yes: the AI Agent node, the model, tools and memory all connect in the visual editor. Code (JavaScript or Python) becomes useful to transform data or call an API that has no dedicated node.
Which AI model should an n8n agent use?
n8n connects to the main providers, including OpenAI, Anthropic and Google, and can combine several models in one workflow. The right choice depends on the task: a fast, cheap model to sort, a stronger one to write or reason. Our comparison: which AI to choose.
Can an n8n AI agent act without human approval?
Yes, but we advise against it for irreversible actions. n8n lets you require human approval on selected tools only: the workflow pauses, a request goes to Slack, Teams, Telegram or n8n's chat, and the action only runs once approved.
What is the difference between an AI agent and a classic n8n automation?
An automation follows a path set in advance. An agent chooses its actions based on the situation. We often combine both: automation for the predictable, the agent for what requires reading, understanding or deciding.
How much does an AI agent built with n8n cost?
Three items: n8n itself (free when self-hosted, from €20/month in the cloud), AI model calls billed by the provider, and the design work. Ballpark figures are in how much does an AI agent cost.
Does n8n support MCP?
Yes, both ways. An n8n agent can call external MCP servers, a workflow can itself become an MCP server, and since version 2.13 you can create or edit workflows by describing them to a tool like Claude.