The short answer
n8n is a workflow automation platform from n8n GmbH in Berlin that you can use on n8n Cloud or host yourself for free with the Community Edition. Its AI features are the AI Agent node, n8n Assistant, introduced on 9 September 2026, which builds and fixes workflows from a description, and n8n Agents, introduced on 25 September 2026, which you reach in Slack, Telegram, Linear or Discord, on a schedule or from workflows. Agents work with any model you have credentials for.
| Made by | n8n GmbH, Germany |
|---|---|
| Free plan | Free self-hosted Community Edition; cloud free trial with 1,000 executions |
| Models | Any model you have credentials for, or n8n Gateway credits |
| Best at | AI agent workflows with full control, code and self-hosting |
| Shows sources | Not applicable; execution logs show which model answered and every step |
| Team plans | Business (self-hosted) and Enterprise, with SSO and environments |
| Your data | Assistant data is not used to train AI models; credentials and workflow data are not sent |
What n8n is now
n8n builds workflows as connected nodes on a canvas, like Zapier or Make, but it leans technical. You can drop in a Code node, call any API, and run the whole thing on n8n Cloud or on your own server. For AI work it has three layers, and knowing which to use is most of the skill.
- The AI Agent node sits inside a workflow. Every AI Agent node now works as a Tools Agent: you attach a model, memory and tools, and it decides which tools to call. Since n8n 2.40 you can force a tool call on the first iteration.
- n8n Assistant builds workflows for you. It launched on 9 September 2026, replacing the earlier AI Workflow Builder.
- n8n Agents, launched on 25 September 2026, are standalone agents you set up once and reach from chat apps, a schedule or other workflows. They do not replace the AI Agent node, and existing AI Agent workflows keep working.
A big release is close. n8n 3.0 has been in beta since 27 October 2026, with the stable version expected in early December. The upgrade is one-way, so back up first, and it removes older AI Agent modes and the legacy OpenAI node.
n8n Assistant: describe it, then check it
Tell the Assistant what you want in plain language. It plans the workflow, builds it on your canvas, asks for credentials when needed, runs it, diagnoses failures and applies fixes. It asks before credential access and before switching a workflow on. The result is a standard workflow you can read and edit, which matters: nothing is hidden in a black box.
Since n8n 2.43 on 6 October 2026, Browser Use is available to all Assistant users. In the 3.0 release notes the Assistant lost its Preview label. On n8n Cloud it is on by default, though Enterprise Cloud is excluded for now. Self-hosters can run it on Docker with version 2.36 or later, their own API keys and some extra environment variables. Builds use Assistant credits, and long debugging rounds use more. The Starter cloud plan includes 1,600 Assistant credits a month.
n8n Agents: one agent, many doors
Open the Agents tab and click Create Agent, or describe the agent to n8n Assistant and let it draft one. You write the goal, pick a model, and give it tools and workflows. The same agent can then answer in Slack, Telegram, Linear or Discord, run on a schedule, or be called from any workflow with the Message an Agent node.
Two controls make this safe enough for real work. Mark a tool as sensitive and the agent pauses until someone approves or rejects the action; n8n recommends this for anything that writes to a system of record. Each tool also uses its own credential, so the agent never holds your instance keys. Skills, which are reusable instructions and reference files, load only when a task needs them and can be shared across agents.
The pricing is generous for agents: one turn with an agent counts as one execution, and calls to tools, workflows and sub-agents inside that turn do not count separately. Agents are still in preview, and Enterprise access is in preview too.
Concrete automations you can build
Support agent with approvals
n8n's own example: an agent reads new tickets, pulls account context, drafts replies, logs notes and escalates urgent issues. Mark the tool that pages on-call as sensitive so a person approves it.
Inbound lead researcher
A workflow triggered by a form sends each lead to an agent with a web search tool. The agent researches the company and posts a short summary to Slack.
Morning summary
An agent on a schedule reports what changed overnight, such as new issues in Linear, failed workflows or new sign-ups, in one Slack message before the team starts.
Internal Slack helper
Connect an agent to one or two systems your team asks about all the time, such as your wiki and your ticket tracker, and let people ask it in Slack. Start small; n8n suggests one or two systems.
Where n8n falls short
- It assumes some technical comfort. Expressions, JSON and credentials setup will slow non-technical users down.
- Self-hosting is work. You own updates, backups and security, and the self-hosted Assistant needs manual setup.
- Things are moving fast. Agents are in preview and 3.0 brings breaking changes, so pin your version and read release notes before upgrading.
- No permanent free cloud plan. Cloud starts with a free trial; the free option is the self-hosted Community Edition.
How to get better results
- Name the trigger, the apps and the output. The Assistant asks fewer clarifying questions when you do.
- Review the built workflow before activating it. n8n's own team says generated workflows should be checked before production.
- Mark every write action as sensitive until you trust the agent, then relax approvals one tool at a time.
- Use the agent budget settings, added in 2.43, so a looping agent cannot run up costs.
- Read the logs. Since 2.45 they show which model answered, which helps when you compare models.
Privacy and which plan suits whom
n8n's privacy documentation says the Assistant sends AI services only context about the current workflow: which nodes exist, data schemas without values, node settings and code. It does not send credential values or the actual data your workflows process, and that data is not used to train the AI services' models. Never paste secrets into the chat; enter them only through n8n's credential screens. On self-hosted instances, admins can switch off telemetry with N8N_DIAGNOSTICS_ENABLED=false, and stop version checks and template calls with N8N_VERSION_NOTIFICATIONS_ENABLED=false and N8N_TEMPLATES_ENABLED=false. Telemetry cannot currently be switched off on n8n Cloud. Self-hosting is the strongest privacy option, because workflow data stays on your own servers.
Starter suits a small team trying AI workflows, with 2,500 executions a month. Pro raises that to 10,000 with more Assistant credits. Business, currently self-hosted only, adds SSO and environments, and companies under 20 employees get 50% off. Enterprise adds 200+ concurrent executions and long insight history. See the n8n pricing page for current prices.
Pick n8n if you have someone technical and want control over models, hosting and cost. If not, our Zapier guide or Make guide is the better start; compare them on Zapier vs n8n and Make vs n8n.
Prompts to copy
Ask n8n Assistant for a lead workflow
Build a workflow: when a new submission arrives from [form tool], research the company website, write a three-line summary and a fit score from 1 to 5 against this ideal customer: [description], then post it to the #leads Slack channel. Ask me for credentials when you need them and do not activate the workflow until I confirm.
You get: A working, editable lead workflow with nothing switched on yet
Instructions for a support agent
You are our support agent. For each new ticket: look up the customer's account, draft a reply under 150 words in plain, friendly English, and add an internal note with what you checked. Escalate to a human if the customer mentions cancelling, a security issue or a payment failure. Never promise refunds.
You get: Clear agent instructions with escalation rules
Morning summary agent
Every weekday at 8:00, check what changed since yesterday at 18:00: new Linear issues, failed workflow executions and new sign-ups in [app]. Post one Slack message with three short sections and link each item. If nothing changed in a section, say 'nothing new'.
You get: A daily status post the team can read in a minute
Debug a failing workflow
This workflow failed on its last three executions. Find the node where it breaks, explain the cause in plain English and propose the smallest fix. Show me the change before you apply it.
You get: A diagnosis and a fix you approve
Classify with structured output in the AI Agent node
Read the email below and return JSON with these keys only: category (sales, support, invoice, other), urgency (1 to 5), summary (one sentence), needs_reply (true or false). If you are unsure, use 'other' and urgency 3. [email text]
You get: Predictable JSON that a Switch node can route
Internal Slack helper skill
You answer questions from our team in Slack using only our wiki and ticket tracker tools. Quote the page or ticket you used. If neither has the answer, say you do not know and suggest who to ask from this list: [names and areas].
You get: A helper that cites its sources and admits gaps