n8n builds visual workflows with AI, app connections and code steps. Choose hosted or self-hosted operation, then test data handling, failures and approvals before scaling.
n8n
Explore features, practical uses and pricing below.
n8n is a workflow automation platform for connecting applications, moving and transforming data, and adding AI processing to a defined business process. Its visual builder represents a workflow as connected nodes, while expressions, API requests and code steps provide more technical control when the built-in settings are insufficient. It is useful for teams that want to inspect how information moves from an incoming event to a resulting action.
An n8n workflow can be a straightforward integration, such as synchronizing records between systems, or an AI workflow that extracts information, retrieves documents and prepares a response. The platform is not itself one language model. It connects model providers and other services to the rest of the process. That distinction matters because the quality, permissions and costs of the finished automation depend on every connected component.
The workflow documentation describes a workflow as connected nodes that automate a process. Triggers start work, and subsequent nodes process information or call services. This makes the workflow's structure visible: you can see where a request enters, where a condition is checked and where an external action occurs. A diagram can be useful both for building and for discussing the process with colleagues.
For example, a new customer form might trigger a workflow that validates an email address, looks for an existing contact and creates a record only when one is missing. Those steps express different responsibilities. Separating them makes it easier to investigate a failure than combining everything into one opaque action. Give nodes descriptive names based on their role, especially when several nodes call the same service for different reasons.
Data inspection and mapping let builders look at node inputs and outputs, reference earlier values and transform records. n8n documents table, JSON and schema views, along with transformation nodes and expressions. This is useful when two systems use different field names, date formats or structures. Check the actual dataset rather than assuming the first example represents every item the workflow will receive.
A practical data contract should say which fields are required, how missing values are handled and which identifier connects records across systems. If one service calls a person a contact and another calls them a customer, an email address may seem like the obvious matching key. It may still be editable, shared or absent. Decide how your process should handle that case before relying on it for updates.
n8n's flow logic includes conditional branches, merging, loops, waits and sub-workflows. A workflow can route records differently based on a condition or delegate a reusable portion to another workflow. This is helpful when an ordinary process contains exceptions: an incomplete lead may go to a review queue, while a complete one is sent to the appropriate sales system.
The HTTP Request node can call REST APIs and supports request configuration such as methods, authentication, query parameters, bodies and pagination. It can be used as a workflow step or as a tool attached to an AI agent. This gives technical builders a route to services or operations that are not fully represented by a ready-made app node.
Using an API still requires understanding its contract. Confirm the endpoint, required scope, pagination behavior and response shape against the service's official documentation. A request that succeeds for the first ten records may silently omit the rest if pagination is wrong. A successful HTTP response can also contain an application-level failure. Make the workflow inspect the information that establishes the business result, rather than stopping at a green node.
Code and expressions support custom logic, including JavaScript or Python in the Code node, subject to the documented runtime and environment. A short expression may be sufficient to normalize a value; a code step can handle a transformation that would otherwise require many visual steps. Keep complicated logic documented and reviewable. When a large code step becomes the core of a business system, consider whether a separately maintained service would make the dependency easier to operate.
The AI integration guide covers model providers, agents, tools, memory and related components. A workflow might call a model to summarize a document, classify a message or extract specific fields. Those model outputs can then be checked and routed through ordinary automation steps. AI should have a defined role in the process rather than becoming the explanation for every decision.
n8n distinguishes chains from agents. A chain follows a configured sequence of AI components, while an agent uses a model to decide which actions or tools to use. The documentation also notes that n8n's chain nodes do not retain conversational memory. Choose the simpler fixed flow when the task already has known steps; use an agent when choosing among tools is part of the problem.
For a support classifier, a fixed flow may be sufficient: receive the message, ask for a category from a defined set, validate the response and assign the ticket. An agent could be more relevant when the answer requires deciding whether to inspect an order, retrieve policy information or ask for missing details. Tool access should correspond to that purpose. A read-only order lookup does not require the same authority as issuing a refund.
Retrieval-augmented generation connects model responses with external documents using components such as data loaders, embeddings and vector stores. This can support a knowledge assistant that retrieves relevant help content before drafting an answer. Plan how documents are added, updated and removed. Retrieved context can be outdated, irrelevant or unavailable to the requesting user, so a retrieval step does not automatically guarantee an accurate or appropriately authorized answer.
n8n provides human review for selected agent tools. When configured, the workflow pauses before a tool action and presents its proposed parameters for approval or denial. The documentation describes review channels including n8n Chat and supported communication services. This makes oversight part of the workflow rather than depending only on an instruction telling the model to be careful.
A useful approval request should show the action, the target and the information being changed. “Approve update” provides little basis for a decision. Showing the customer, current status, proposed status and reason lets the reviewer understand the consequence. Decide what happens when approval is denied or never arrives, and make that outcome visible to the person waiting for the workflow.
The evaluation documentation describes running datasets through an AI workflow and comparing outputs. A representative set is more informative than a single successful demo. For message classification, include clear examples, ambiguous messages, missing context and text that tries to instruct the classifier to ignore its task. Assess whether the workflow produces the required category and whether its downstream action remains appropriate.
As an editorial recommendation, track several outcomes separately. Correct extraction, appropriate routing and successful delivery are different results. A model can extract the right order number while the integration updates the wrong ticket. A service can accept a request while the generated response is unhelpful. Separating those observations makes it easier to locate the real source of a problem.
Consider a hypothetical online store receiving requests about delivery, returns and product information. An initial n8n workflow could suggest a category, look up limited order information and prepare a response for a support person. It would not autonomously change orders or send customer messages in this first version. The following is an original example workflow, not a claim of hands-on testing or a prediction of accuracy.
After the first version works, add one new capability at a time. Order changes and refunds need their own acceptance criteria, permissions and approval boundaries. Keeping them separate makes the expanding automation understandable to the people who operate it.
Error workflows can respond when a workflow execution fails. n8n documents an Error Trigger, associated execution information and reviewing prior runs. Configure the failure response early, even for a small workflow. A process that stops quietly can leave a team believing an invoice, notification or record has been created when it has not.
Failure handling should distinguish a temporary problem from invalid input. A timeout may deserve a controlled retry; a malformed record may need correction. Repeating a create action can produce duplicates if the earlier call succeeded but its response was lost. Use stable identifiers and the external service's supported duplicate-prevention behavior where available. Do not assume that rerunning a visually simple workflow is free of side effects.
Keep logs useful without making them a second unmanaged copy of sensitive data. Decide what execution information is saved, who may read it and how long it is retained. A workflow operator should be able to identify the failing step and affected record without exposing every credential or entire customer document. Review retention and collaboration settings for the chosen plan and edition.
n8n Cloud is the managed hosting option. Self-hosting places the software on infrastructure you operate, with Community and paid editions available. These are two separate decisions: who operates the infrastructure, and which features or limits you need. An organization can require paid governance features even when it chooses to run its own server.
The hosting documentation covers deployment and maintenance. Self-hosting requires capacity to manage persistent storage, access, updates, monitoring and recovery. Assign those responsibilities before making a workflow business critical. A server that starts successfully is not evidence that backups can be restored or that an upgrade will preserve every integration.
n8n's Community licensing guide explains the Sustainable Use License and the fair-code model. The source is available, but n8n does not describe this as an unrestricted open-source license. Its license FAQ distinguishes operating workflows internally from offering customers the ability to build or configure workflows. Check those terms against the actual product architecture if you intend to embed, resell or provide hosted access.
Self-hosting also does not keep every workflow interaction inside your infrastructure. A node calling a cloud model, email provider or other external service sends data through that service according to its configuration. Map those connections and use only the information the task needs. Hosting control and third-party data processing are related but separate considerations.
At the October 8, 2026 research check, n8n's pricing page listed Cloud Starter at €20 per month billed annually with 2,500 monthly workflow executions, and Pro from €50 per month billed annually with 10,000 executions. It also listed a paid self-hosted Business plan and custom Enterprise options. An execution is a workflow run rather than a separate charge for each node. Free Cloud trials and the self-hosted Community edition provide different routes to trying the product.
Estimate usage from real triggers: how often a schedule runs, how many webhook events arrive and how many messages start an AI flow. Compare concurrency, execution duration, saved history and administrative requirements alongside the execution allowance. External API subscriptions, model calls, storage and self-hosted infrastructure may add costs beyond the n8n plan.
Current Gateway credits let supported Cloud plans run eligible models and services without their own provider credentials. The documentation limits this option to specified Cloud plans and services, and distinguishes it from Assistant credits. Workflows can also use their own provider credentials. Check the supported operations and balance behavior for your setup; model access is not an unlimited free component of every workflow.
Many steps can be configured visually. Technical knowledge becomes useful when mapping nested data, calling APIs or handling failures and credentials. Start with one process you understand, then add code only where it makes that process clearer or more capable.
No. Use ordinary logic for known conditions and fixed steps. An agent is relevant when choosing tools or deciding a sequence is part of the task. A straightforward synchronization often needs explicit mappings and error handling more than open-ended reasoning.
The self-hosted Community edition can be used under its license terms, while hosting and connected services may still cost money. Cloud access has trial and paid-plan conditions. Check both feature requirements and operational responsibilities before treating either route as equivalent.
Use the same process when comparing Zapier, Make or Gumloop. Check the exact integrations, data transformations, failure recovery, human review and ongoing cost. For n8n, also decide whether controlling the deployment is valuable enough to justify operating it yourself.