n8n Review: Self-Hosted Automation, AI Agents, MCP and Scale

n8n's current AI architecture is becoming less about placing a chatbot inside an automation and more about treating model behavior as one controlled component of a production workflow.

AI Agents can call application nodes, custom tools, subworkflows and external MCP servers. Sensitive tool calls can pause for human approval, while fallback branches can transfer an unresolved task to a person or deterministic process. Guardrails can inspect model inputs and outputs before the result reaches another system.

Evaluations address a different production problem. A workflow can be run against a test dataset and scored using expected outputs or defined metrics. This allows a team to compare prompts, models, retrieval strategies and workflow versions rather than approving an agent from a few successful demonstrations.

MCP works in both directions. An n8n agent can use tools from an external MCP server. An n8n workflow can also be exposed as an MCP service, allowing compatible assistants to call a constrained business process instead of receiving broad direct access to an underlying application.

These controls improve accountability but do not make an agent safe automatically. The team still needs representative evaluation data, scoped credentials, explicit approval policy, timeouts, rejection handling, audit evidence, cost monitoring and a deterministic recovery path when the model or an external tool fails.

WhatAI Decision Box

n8n is a fair-code workflow automation and AI orchestration platform for technical teams. It combines a visual node editor, JavaScript and Python, self-hosting, application integrations, AI agents, RAG, evaluations, human approval, MCP, production debugging, queues, Git environments and enterprise governance.

n8n gives technical teams more control over code, deployment, data location, custom nodes, APIs and AI architecture. Hosted pricing counts a complete workflow run as one execution regardless of the number of steps. Self-hosting can remove a Cloud execution quota, but it replaces subscription simplicity with infrastructure and security responsibility.

A visual canvas does not remove software-operating risk. Production workflows need versioning, validation, retries, idempotency, observability, scoped credentials, tested recovery, controlled community nodes, backups and owners. AI agents additionally need evaluations, guardrails, human approval and clear limits on tool access.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-07-24.

n8n Community Discussions

Explore community discussions. Ask and answer questions on n8n to grow and learn together.

n8n Pros & Cons

How to Get Results with n8n: Step-by-Step Workflow

1. Define the business and failure outcome

   State the trigger, required result, systems, owner, expected volume, latency, data sensitivity, acceptable failure, manual fallback and evidence that one execution succeeded.

   Decision point: Use deterministic nodes for known rules. Add an AI Agent only when interpretation, retrieval or adaptive tool choice is genuinely required.

2. Choose Cloud or self-hosting

   Compare execution volume, data location, network access, integration requirements, engineering capacity, support, security obligations and commercial features before selecting deployment.

   Decision point: Choose Community only when the team can operate and secure the service and the intended use complies with the license.

3. Design the workflow contract

   Define trigger payloads, required fields, schemas, idempotency keys, outputs, error classes, timeouts, retries, credentials, retention and downstream side effects.

4. Build the smallest deterministic path

   Connect the trigger, validate the input and produce one verified outcome before adding branches, loops, AI, code or integrations.

5. Modularize complex logic

   Move reusable or independently owned capabilities into subworkflows with explicit inputs and outputs.

   Decision point: Split the graph when another engineer cannot understand the happy path, failure path and ownership quickly.

6. Add AI with bounded tools

   Choose the model, prompt, retrieval, memory, output schema and minimum tool set. Keep irreversible actions outside the agent or behind explicit approval.

7. Create evaluation data

   Collect representative normal, ambiguous, adversarial and failure cases. Define expected outputs and metrics before changing prompts or models.

8. Design approval and fallback

   Specify which tool calls require a person, who receives the request, timeout behavior, rejection handling, escalation and the deterministic fallback when no decision is received.

9. Add production error handling

   Classify temporary, permanent, authentication, rate-limit, duplicate, malformed-data and partial-side-effect failures. Add capped retries, dead-letter storage, alerts and replay.

10. Secure and deploy

    Use scoped credentials, encryption, private networking, secret management, approved nodes, protected environments, backups, patching, security audit and least-privilege roles.

11. Monitor real operation

    Track executions, queue depth, worker health, latency, failures, retries, model cost, approval delay, tool calls, operator interventions and the actual business result.

12. Version and improve

    Review evaluation results, incidents, execution samples and cost. Update the workflow through a controlled environment and preserve rollback, ownership and change evidence.

n8n Gotchas and Limits to Know Before You Start

• n8n describes itself as fair-code and the Community Edition uses the Sustainable Use License
• Source availability does not grant unrestricted resale hosted-service or embedding rights
• Community Edition has no n8n subscription fee but infrastructure and operating costs remain
• Community Edition excludes several commercial collaboration governance storage and enterprise capabilities
• Starter and Pro public prices are displayed in euros at the current annual-billing rate
• Business is currently positioned as a self-hosted commercial plan at a substantially higher annual price
• One workflow run is one paid execution regardless of node count
• Each polling interval webhook event scheduled run or chat message can create another execution
• AI Assistant credits are separate from workflow-execution quotas
• Unused AI Assistant credits currently do not roll over
• Self-hosted AI model and provider costs are separate from n8n licensing
• Queue mode requires Redis and a supported database architecture
• Production worker scaling requires concurrency queue health and backpressure planning
• Binary files require a storage strategy that works across workers
• External S3 binary storage is an Enterprise capability
• Task runners should be used to isolate untrusted or resource-intensive Code-node execution
• Execute Command community and custom nodes can run code with host-level security implications
• Community nodes are third-party packages and need source dependency and permission review
• The security audit identifies risks but does not replace hardening penetration testing or monitoring
• Git source control does not move live credential values into the repository
• Git push and pull design can overwrite work when environments and branches are used carelessly
• Business and Enterprise environments require a deliberate development and production promotion process
• AI Agent Tool and subworkflow tools can create deep call chains that are hard to trace
• An AI agent may select the wrong tool or supply incorrect parameters
• Human approval can time out or become an operational bottleneck
• Evaluations are only as useful as the cases labels and metrics in the dataset
• Passing an evaluation set does not prove safety on unseen production inputs
• MCP Server Trigger can expose consequential workflows to external AI clients
• MCP Client can inherit security availability and schema risks from external servers
• Memory can retain irrelevant or sensitive context unless scope and retention are controlled
• RAG can retrieve outdated irrelevant or unauthorized documents
• Large workflows need naming notes subworkflows owners and architecture documentation
• Execution retention limits can remove debugging evidence on hosted plans
• Deleted workflows can remove associated execution history
• Self-hosted upgrades require database backups compatibility review and rollback planning
• Enterprise license keys communicate usage information to n8n under current licensing behavior

Which n8n Feature Fits Your Use Case

How Well n8n Fits Common Use Cases

Starter Prompts for n8n

Design an n8n workflow for [business outcome]. Define the trigger, input schema, idempotency key, validation, application nodes, API calls, transformations, subworkflows, success output, temporary and permanent failures, retry limits, dead-letter handling, alerts, credentials, retention, owner and production acceptance tests.

Design an n8n AI Agent for [task]. Keep deterministic rules outside the agent. Define the model, prompt, memory, retrieval, structured output, approved tools, prohibited actions, human-approval tools, timeout, rejection flow, fallback workflow, evaluation dataset, metrics, logging, model cost and escalation owner.

Design an n8n workflow exposed through MCP for [business outcome]. Specify authentication, input schema, validation, authorization, idempotency, rate limits, scoped credentials, approval, deterministic execution, clean output, error codes, audit evidence and the actions the external AI client must never control directly.

Design a production self-hosted n8n architecture for [expected executions and data sensitivity]. Cover database, Redis, queue mode, workers, task runners, binary storage, encryption key, secrets, network controls, TLS, backups, monitoring, logging, upgrades, rollback, disaster recovery, capacity testing and license requirements.

Create an n8n Evaluation plan for this AI workflow. Build a representative dataset with normal, ambiguous, adversarial, policy, retrieval and tool-failure cases. Define expected output, factuality, format, tool-choice, safety, latency and cost metrics. Explain thresholds, regression rules and manual review sampling.

Prompt pattern: When [trigger] occurs, validate [schema and identity], then execute [deterministic steps]. Allow AI to [bounded task] using only [tools and data]. Require approval for [sensitive actions]. Handle [failure classes]. Return [structured output]. Log [audit evidence metrics and cost].

Iteration tip: Stabilize the deterministic path before adding an AI Agent. Then add one tool, one approval policy and one evaluation dataset at a time so failures remain attributable.

What is n8n best used for in 2026?

n8n is best used for technical workflow automation, AI-agent orchestration, custom API integration, self-hosted internal automation, data movement, operational tooling, MCP services and processes that need more logic than simple trigger-and-action products provide.

Is n8n open source?

n8n publishes its source and provides a free self-hosted Community Edition, but its documentation describes the product as fair-code under the Sustainable Use License rather than a conventional open-source license. Certain commercial hosting, resale and embedding uses require separate permission or licensing.

Is self-hosted n8n free?

The standard Community Edition has no n8n subscription fee. The operator still pays for infrastructure and owns installation, upgrades, database maintenance, backups, security, monitoring and support. Commercial features and license restrictions also apply.

How much does n8n Cloud cost?

At the current annual-billing prices, Starter costs twenty euros per month for two thousand five hundred executions, and Pro costs fifty euros per month for ten thousand executions. Business is a self-hosted commercial plan at six hundred sixty-seven euros per month billed annually for forty thousand executions. Enterprise is custom.

How does n8n count executions?

On paid plans, one complete run of a workflow is one execution regardless of how many nodes or steps it contains. A chatbot can still consume many executions because each incoming message or triggered conversation turn may start another workflow run.

Can n8n build AI agents?

Yes. n8n supports AI Agent nodes, tool calling, subworkflow tools, model selection, memory, retrieval, vector stores, reranking, structured output, guardrails, human fallback, human approval and evaluations.

What MCP features does n8n provide?

n8n can connect to external MCP servers through MCP Client nodes, give MCP tools to an AI Agent through MCP Client Tool, expose a workflow through MCP Server Trigger and provide an instance-level MCP server for supported workflows and n8n resources.

Can a human approve an n8n AI agent action?

Yes. Supported tool calls can pause and request human approval through configured communication nodes before the agent executes a sensitive action. Builders should still design timeout, rejection, escalation and audit behavior.

How do n8n Evaluations work?

Evaluations run AI workflows against test datasets so teams can compare output with expected results or score it using defined metrics. They help detect regression, but the dataset and metrics must represent real production risk.

Is n8n easier than Make or Zapier?

Usually not for a complete beginner. n8n gives technical users more control over code, APIs, self-hosting, AI and deployment, while Make and Zapier generally provide a more guided managed experience for straightforward business automation.

Related Automation & Process Tools

Explore the Network