Guides to n8n and AI agents.

The n8n AI Agent is a node in which a language model receives instructions (a system message) and a set of tools, then decides on its own which tools to call to complete a task. Without tools it is just a chat. With tools, most often sub-workflows, the agent can read a document, check data in a spreadsheet and make a judgement call no plain rule can handle. Since August 2026 n8n also has a separate Agents category: agents with their own memory, channels and skills that live outside a single workflow.

Claude Code connects to n8n through the n8n-mcp MCP server: one terminal command adds the server, and the agent gets documentation for every n8n node, configuration validation and workflow operations on your instance. From then on it can build, fix and test a workflow from a description in plain language. More than 143,000 people use n8n-mcp, and its telemetry shows agents spend four times more calls reading and checking than writing.

n8n Cloud starts at €20 a month for 2,500 workflow executions, and the price grows with executions. Self-hosted n8n in the free Community Edition has no execution limit, costs what the server and maintenance time cost, and keeps data inside the company. For companies handling documents, personal data or larger volumes I recommend your own server. Cloud makes sense at the start, when nobody in the company wants to run a server.

n8n-skills is a set of 14 skills plus a skill router and a hooks layer that teach an AI agent the right patterns for building n8n workflows through the n8n-mcp server. You install it with one command in Claude Code, as zip files in Claude.ai and ChatGPT, or as a plugin in Codex. An agent without skills guesses at syntax and puts instructions in the wrong place; an agent with skills sticks to proven patterns and conventions.

Most n8n failures fall into three groups: an agent or client can't connect to the instance, a node that used to work suddenly breaks, or the workflow goes silent and nobody finds out. Executions shows you what happened, but not what never ran. Without a separate error workflow, silence is the default state: nobody checks executions every day, so a failure waits until someone happens to open the tab.

Zapier counts tasks per successful step, Make counts operations per module, n8n counts executions regardless of how many steps a workflow has. That billing unit matters more to the bill than the integration list on the homepage. Zapier and Make win on onboarding speed and the number of ready integrations, n8n wins on volume, data control, and code where the ready-made blocks run out.

An off-the-shelf AI receptionist starts at $29 a month and can be live in a few hours. A custom voice agent on GPT-Live with Twilio telephony runs at about $0.065 per minute plus roughly $14 in fixed monthly fees, but someone has to build it. Below about 50 calls a month the ready-made service wins; above 500 a custom agent pays for itself within months. In both cases bookings, CRM lookups and quotes run in n8n, which the agent calls mid-conversation.