Build an AI Agent in n8n

By the end of this tutorial you will have a chat agent running in your own n8n instance. It answers questions through a chat window, remembers the last few messages, and uses two tools, Wikipedia and a calculator, to look things up and do sums. The shortest path is four nodes on a canvas: Chat Trigger, AI Agent, a chat model and a memory node, plus tools. Allow an hour. Facts checked 31 October 2026.

What you will have at the end

• n8n running locally in Docker, with your data in a volume
• A workflow with a Chat Trigger, an AI Agent, a chat model, Simple Memory and two tools
• A tested agent, with questions that show which tool it used
• A plan for human approval on any tool that acts in the real world
• A cost picture for self-hosting and for n8n's cloud plans

If you would rather write the loop yourself, see how to build an AI agent. If you want a visual builder, read on. For the product itself, see n8n and n8n pricing.

A note on alternatives. OpenAI's Agent Builder shuts down on 2026-11-30, so do not start a new project there. See OpenAI Agent Builder and the list of AI agent platforms.

Prerequisites

• Docker Desktop, or Docker Engine on Linux. The n8n docs recommend Docker for most self-hosting needs.
• An API key for a chat model provider. This guide uses Anthropic's chat model node, but n8n has nodes for many providers.
• A modern browser

The n8n docs add a warning worth reading: self-hosting requires knowledge of servers, containers, scaling and security, and mistakes can lead to data loss, security issues and downtime. n8n recommends n8n Cloud if you are not experienced at managing servers. For a local tutorial, Docker on your laptop is fine. Do not expose it to the internet yet.

Step 1: Start n8n in Docker

The docs give this sequence. Replace both timezone placeholders with yours, for example Europe/Berlin.

docker volume create n8n_data
docker run -it --rm \
 --name n8n \
 -p 5678:5678 \
 -e GENERIC_TIMEZONE="<YOUR_TIMEZONE>" \
 -e TZ="<YOUR_TIMEZONE>" \
 -e N8N_ENFORCE_SETTINGS_FILE_PERMISSIONS=true \
 -v n8n_data:/home/node/.n8n \
 n8nio/n8n

What it does, according to the docs: it exposes port 5678, sets the timezone for the container and for schedule nodes, enforces secure permissions on the settings file, and mounts the n8n_data volume at /home/node/.n8n, so your workflows and credentials survive a restart. Open http://localhost:5678 and create your owner account.

Two things to know:

1. The --rm flag removes the container when it stops. Your data stays in the volume. Never delete the volume unless you mean to.
2. n8n also publishes an npx n8n shortcut, but the npm install docs say npm-based installs are deprecated from n8n 3.0 and that the n8n Assistant is not available with them. Use Docker.

By default n8n stores data in SQLite. It also supports PostgreSQL through environment variables, which you should consider before any production use. The docs also note that the .n8n directory still matters with PostgreSQL, because it holds encryption keys and other data.

Step 2: Add the Chat Trigger

Create a new workflow. Add the Chat Trigger node. The docs describe it as the node to use when building AI workflows for chatbots and other chat interfaces. It has options you will see later:

• Make Chat Publicly Available. Off means the chat works only through the manual chat interface in the editor. Leave it off for now.
• Hosted Chat. This is n8n's own interface, recommended for most users because you can configure it from the node options.
• Authentication. None means anyone with the link can use the chat. Basic Auth adds a login.

For this tutorial, keep the chat private and use the built-in chat panel in the editor.

Step 3: Add the AI Agent node

Add the AI Agent node and connect the Chat Trigger to it. The docs describe it as the node where you connect a chat model and one or more tools, and the agent decides which tools to call to complete a task.

Three facts from the docs matter:

• You must connect at least one tool sub-node to an AI Agent node.
• The agent type setting is deprecated from n8n 1.82.0. All AI Agent nodes now work as a Tools Agent.
• Version 1 of the node, the one with the agent type setting, will be removed from n8n 3.0. If a template you find uses it, update the node.

So there is nothing to choose between agent types on a current install. The node is a Tools Agent.

Step 4: Connect a chat model

On the AI Agent node, click the chat model connector and add a model node. For Anthropic, that is the Anthropic Chat Model. Create a credential with your API key, then pick a model. Store the key in n8n's credential system, never in a node field or a note on the canvas.

Which model? For a tool-using agent, start with a mid-size model, and move up only if it fails on your own test questions. Our Claude models page and LLM API prices help you compare. If your data cannot leave your network, n8n also has nodes for local models, and our guide on how to self-host an LLM shows how to run one. Ollama's docs list n8n as an integration.

Step 5: Add memory

Add the Simple Memory node under the agent's memory connector. It keeps recent chat history so the agent can follow a conversation. Two parameters from the docs:

• Session Key. The key used to store the memory in the workflow data.
• Context Window Length. How many previous interactions to consider.

Set the window to a small number, such as 5, to start. A longer window carries more context and costs more per message, because the history is re-sent each turn. Note the limit in the docs: do not use Simple Memory if your n8n instance runs in queue mode. In an active production workflow, n8n cannot guarantee that every call goes to the same worker. For queue mode, use a database-backed memory node instead, such as Postgres or Redis chat memory.

Step 6: Connect tools

Tools are what make it an agent. Add two sub-nodes to the tools connector:

1. Wikipedia. Lets the agent look up facts.
2. Calculator. Lets it do exact arithmetic, instead of guessing.

n8n has many more: HTTP-based tools, a Custom Code Tool, the Call n8n Workflow Tool, an MCP Client Tool, search tools such as SearXNG and SerpApi, and a Think Tool. Add them only when you can say what each one is for. See MCP servers and how to build an MCP server if you want to wire in your own tools.

Many tool parameters can be filled in by the model itself. The docs describe a star-icon button on each parameter and a $fromAI() function for this. Use it for free-text inputs like a search query. Do not let the model choose values that matter, such as an email recipient or an amount, without a human check.

Step 7: Give the agent clear instructions

In the AI Agent node's options, set a system message. Keep it short and concrete.

You are a research helper. Answer questions using the Wikipedia tool for facts and the Calculator tool for any arithmetic. Always say which tool you used. If the tools do not give you an answer, say you could not find it. Do not guess numbers.

Step 8: Test it

Open the chat panel in the editor and try questions that exercise each tool.

Check the executions view to see the order of calls and the data each tool returned. That is how you debug an agent: read what it called and what came back. Run each question more than once, because model output varies.

Step 9: Add human review before anything risky

This agent is harmless: it reads and calculates. The moment you add a tool that sends an email, edits a record or spends money, add a person to the loop. n8n supports human review for tools.

How it works, per the docs: when the agent decides to use a tool with review turned on, the workflow pauses and sends an approval request through the channel you configure, such as Slack, Telegram or the n8n Chat interface. The reviewer sees which tool the agent wants to use and with what parameters, and approves or denies it. If approved, the tool runs with the model's input. If denied, the action is cancelled and the agent is told. To set it up, click the Tools connector on the AI Agent node, find the Human review section and choose a channel and its credentials. You can apply review to all tools or to selected ones.

The docs' advice is sound: start with review on, then reduce oversight as you build trust. Keep it on permanently for irreversible actions.

Step 10: Move toward production, carefully

When the agent works in the editor, you can publish the chat. Before you do, work through this list.

1. Authentication. Turn on Basic Auth on the Chat Trigger, or put the instance behind your own login. "None" means anyone can use your model credits.
2. Reachability. The Docker command above is for local use. Put production behind HTTPS and a reverse proxy, and read n8n's configuration docs.
3. A real database. Use PostgreSQL rather than the default SQLite for anything serious.
4. Backups. Back up the n8n_data volume and your encryption key.
5. Updates. n8n releases a minor version most weeks. The docs recommend the stable tag for production and say beta may be unstable. Pull the new image and recreate the container to update.
6. Limits. Add a spend cap at your model provider.
7. Tunnel. The docs say n8n's tunnel feature is for local development and testing only and is not safe in production.

What can go wrong

What it costs

Two separate bills: n8n, and your model provider.

n8n. From the pricing page, checked 31 October 2026. All plans include unlimited users and workflows and every integration, and pricing depends on monthly workflow executions.

An "execution" is one full run of a workflow, so a chat message that triggers the agent is one execution. If it does not finish, you do not pay, according to the page.

Model provider. You pay per token for the chat model. The agent loop re-sends history and tool results each step, so cost per message grows with memory and tool use. Use your provider's usage dashboard during testing and write down a per-message figure. See LLM API prices.

Your time. A self-hosted instance needs patching, backups and monitoring.

Privacy

On a self-hosted instance, your workflows, credentials and execution data live on your server, in your volume. Chat text and tool results still go to the model provider you choose, and Wikipedia calls go to Wikipedia. Execution logs can hold user messages, so decide how long to keep them. If you use n8n Cloud instead, n8n hosts that data, so read its privacy and security pages. For regulated data, pair self-hosted n8n with a local model, as in private AI for teams.

Checklist

• n8n running in Docker with a mounted n8n_data volume
• Owner account created, instance not exposed to the internet
• Chat Trigger, AI Agent, chat model, Simple Memory and at least one tool connected
• API key stored as an n8n credential
• A short system message that names tools and forbids guessing
• Five or more test questions run several times, with the execution view read
• Human review turned on for any tool that sends, edits or spends
• Authentication on the chat before it is made public
• PostgreSQL, backups and a spend cap before production
• Monthly execution count estimated against a plan

What this guide uses