n8n is a visual workflow automation platform with cloud and self-hosted deployment options. It combines prebuilt integrations with custom API requests, JavaScript, Python, AI models, memory, and agent tools.
For this review, we assessed its AI architecture, item-based data model, execution billing, cloud limits, and self-hosting requirements.
The AI Agent node is the main attraction. It acts as the decision-making centre of a workflow. You connect a language model, memory, and tools underneath it. Those tools can search data, call an API, run another workflow, or ask a person for approval.
This structure gives the agent boundaries. An email agent might be allowed to classify a message and create a draft. It does not need permission to send the email. That division matters because AI agents can choose the wrong tool or repeat an action when instructions are unclear.
n8n can also work as an MCP client or server. A workflow can use tools exposed by another system. It can also become a tool that an outside AI assistant calls. This is useful for turning a carefully tested business process into one controlled action.
The harder part is n8n’s item-based data model. Every node passes an array of JSON objects to the next node. When one node returns 50 items, the following node may process all 50 automatically.
That behaviour is powerful until an AI node is involved. Passing 50 records into an agent can create 50 separate model calls. Each call has its own token cost and failure risk. The workflow may still count as one n8n execution, but the AI provider can bill for every request.
Avoiding this requires deliberate data handling. You may need to group records before the agent, split them into safe batches, or write code to reshape the JSON. This is the point where the visual canvas stops feeling no-code.
Cloud pricing is based on complete workflow executions. A 20-step workflow counts as one execution when it runs from trigger to finish. That is straightforward for long workflows. Cloud plans also place limits on concurrency. Starter supports five simultaneous executions, while Pro supports 20. Extra runs wait for capacity, so a sudden burst can create delays.
The Community Edition removes the subscription fee and can be self-hosted for internal use. It does not remove the operating work. Someone must secure the server, update the software, back up the database, manage credentials, and watch resource use. Self-hosting makes sense when the team wants control and already knows how to maintain it.
