n8n

n8n is an open-source workflow automation platform built for technical teams who want more control than consumer tools like Zapier offer. You can self-host it for free, write JavaScript or Python anywhere in a workflow, and connect to 500+ services out of the box. Since 2023 n8n has invested heavily in AI: it ships native LLM nodes for OpenAI, Anthropic, Google Gemini, Mistral, and others, plus multiple agent architectures (ReAct, Plan-and-Execute, Tools Agent) and full RAG support with vector store integrations. The visual canvas is genuinely useful for building complex multi-step automations, and the per-step debugging experience is a cut above most competitors. Cloud plans start at €20 per month; self-hosting is entirely free under the Sustainable Use License.

Most automation tools make a quiet bet: that you won’t need to do anything too unusual. n8n rejects that bet. Built for developers and technical teams, it gives you a visual canvas where you can see every step, but it also lets you write real code, connect directly to any API, and now, wire LLMs into your workflows with native AI nodes. If you’ve ever hit a wall in Zapier because there was no “do exactly what I mean” button, n8n is what you were looking for.

Quick verdict

n8n is the right tool if you know what you’re doing and want to build automations that Zapier can’t handle. The self-hosted Community Edition is genuinely free and capable. The AI nodes are flexible enough to build real agent pipelines. The trade-off is that setup takes real effort, and the UI rewards people who think in nodes and JSON. If you’re not technical, look elsewhere.

What is n8n, exactly?

n8n started in 2019 as a simpler answer to a developer frustration: automation tools were either too limited (Zapier) or too heavy (building everything from scratch). Jan Oberhauser launched it as an open-source project, and it grew to over 187,000 GitHub stars with a community of more than 200,000 users. The company behind it, n8n GmbH, is based in Berlin and has taken a deliberate path: keep the core open source, build a managed Cloud product for teams who don’t want to run infrastructure, and push hard into AI automation.

The product is a node-based workflow builder. You drag nodes onto a canvas, connect them, configure them, and run them. That description sounds like Zapier, but the difference is in the details. Every node shows you its input and output data inline. You can re-run a single step without re-running the whole workflow. You can add a Code node anywhere and write JavaScript or Python that runs server-side. You can call any HTTP endpoint, handle webhooks, write conditional branches, loop over arrays, merge data streams, and handle errors with dedicated error-handling nodes.

The canvas gets crowded on complex workflows, which is a real usability problem. But there’s a reason developers keep reaching for it: when something breaks, n8n shows you exactly where and why. The debug experience is meaningfully better than competitors that show you a vague “step failed” message.

In 2023 n8n began releasing dedicated AI nodes, and that work has accelerated. By 2026 the platform supports a full LLM integration stack, multiple agent architectures, vector store connections, and memory management. It now has native MCP support, which matters for connecting to the growing ecosystem of MCP-compatible tools. The GitHub star count has climbed past 187,000, making it one of the most-starred automation projects in the world.

The features that earn it the developer audience

AI Nodes for LLM-powered workflows

The AI integration in n8n isn’t bolted on. It’s built around LangChain concepts, which means the node vocabulary will feel familiar if you’ve worked with chains, agents, or RAG pipelines in Python. You connect an LLM node (OpenAI, Anthropic Claude Sonnet 4.6 or Opus 4.7, Google Gemini 3, Mistral, DeepSeek, Groq, Azure OpenAI) to a Chat Trigger or an HTTP input, give it tools or memory, and you have an agent loop. Vector stores including Pinecone, Qdrant, Chroma, Weaviate, Supabase, and MongoDB Atlas are available as dedicated nodes. Memory can be backed by Redis, Postgres, or MongoDB.

What makes this genuinely useful is that the AI nodes follow the same debugging model as everything else: you see the input going in and the output coming out at each step. When your agent calls a tool, you see the tool call and the result. When memory is retrieved, you see what was retrieved. This level of transparency is harder to get when you’re writing LangChain scripts by hand and piping logs to a terminal.

Code nodes and custom JavaScript

The Code node is the escape hatch that makes n8n viable for real engineering work. At any point in a workflow you can insert a Code node and write JavaScript or Python. The code runs server-side, has access to the full item array passed from the previous node, and can return any structure you want for the next node to consume. There’s also a LangChain Code node for more advanced AI integrations where the visual nodes aren’t granular enough.

This matters because it means n8n is never truly blocked. If there’s no native node for what you need, you write it. If the existing node doesn’t expose the parameter you need, you make an HTTP Request node and construct the call yourself. The escape hatch is always open, which keeps n8n from feeling like a tool you’ve outgrown.

Self-hosting that actually works

Running n8n yourself is not a theoretical option kept alive for marketing purposes. It’s a first-class deployment path with Docker images, Kubernetes helm charts, and guides for AWS, GCP, Azure, DigitalOcean, and Railway. The Community Edition runs on any Linux server with Node.js or Docker. Your data stays on your infrastructure, there are no per-run fees, and you can run as many workflows as your server handles.

The practical overhead is real: you own updates, backups, SSL, and uptime. But for a team with an existing DevOps practice, this isn’t unusual work. Many companies run n8n on a small VPS or in their Kubernetes cluster and treat it like any other internal service. For use cases involving sensitive data, self-hosting removes the question of whether workflow data passes through a third-party cloud.

500+ pre-built integrations

n8n ships native nodes for over 500 services. The list covers the standard SaaS stack (Slack, Google Sheets, Notion, HubSpot, Salesforce, GitHub, Jira, Airtable) plus databases (Postgres, MySQL, MongoDB, Redis), developer tools (HTTP Request, GraphQL, RSS, SSH, FTP), and communication tools (email, SMS, WhatsApp). For anything not on the list, the HTTP Request node handles arbitrary API calls, and you can create custom nodes if the integration is something you need repeatedly.

The integration quality is uneven, as it is on every platform with this many connectors. Popular nodes like Slack, GitHub, and Google Sheets are well-maintained and handle most common actions. Less popular nodes sometimes have gaps or lag behind API changes. The community is active enough that issues usually get filed and patched, but you’ll occasionally need to work around a limitation with a raw HTTP call.

Agent workflows from primitives

n8n ships five distinct agent node types: Conversational Agent, ReAct Agent, Plan-and-Execute Agent, OpenAI Functions Agent, and Tools Agent. These aren’t marketing labels for the same underlying loop. Each has a different decision-making architecture, and choosing the right one for a task genuinely affects reliability and cost. ReAct is the general workhorse. Plan-and-Execute breaks tasks into a plan phase and execution phase, which works better for longer multi-step tasks. The SQL Agent is purpose-built for database queries.

You can also build agent workflows from scratch using just the AI and HTTP Request nodes if the built-in architectures don’t fit your needs. Human-in-the-loop approval nodes let you pause execution, send a Slack message or email asking for a decision, and resume once someone responds. This makes it possible to build agents that handle routine cases automatically and escalate edge cases to a human, which is how most production agent deployments actually need to work.

Pricing

n8n has two cost tracks: self-hosted and Cloud.

Self-hosting the Community Edition is free. You pay for your own server costs, which for a small VPS running a personal or small-team n8n instance might be $5 to $20 per month on DigitalOcean or similar. There are no workflow limits, no execution limits from n8n’s side, and no license fees. Features like SSO, Git version control, and multiple environments require Business or Enterprise tier even on self-hosted, but the core automation and AI features are all available.

Cloud plans are priced in euros and billed annually. The Starter tier is €20 per month and includes 2,500 workflow executions, 5 concurrent executions, 1 shared project, and 50 AI Workflow Builder credits. The Pro tier is €50 per month and raises those limits to 10,000 executions, 20 concurrent executions, 3 shared projects, and 150 AI credits, plus workflow history and admin roles. The Business tier jumps to €667 per month and adds SSO/SAML/LDAP, Git version control, multiple environments, and 40,000 executions. Enterprise is custom-priced.

The gap between Pro (€50) and Business (€667) is steep. Teams that outgrow Pro’s execution limits but don’t need enterprise features will find this pricing awkward. For those teams, self-hosting with a managed database is often the more cost-effective path. The AI Workflow Builder credits let n8n’s built-in AI generate workflows from natural language descriptions, which helps less technical team members create automations without hand-building every node.

Where n8n wins and where it doesn’t

n8n wins on control and depth. If your automation needs custom logic, n8n handles it. If you need to call an obscure API with specific headers, n8n handles it. If you want to run an LLM agent with RAG, memory, and tool calls in a single workflow, n8n handles it. The per-step debugging is genuinely useful when something fails, and the ability to re-run a single node without replaying the entire workflow saves real time during development.

n8n also wins on data privacy for teams that need it. Self-hosting means your workflow data, credentials, and execution history never leave your infrastructure. For companies in regulated industries or with strict data handling requirements, this is a meaningful advantage over cloud-only tools.

Where n8n struggles is the new-user experience. The onboarding assumes a level of technical comfort that many potential users don’t have. The canvas can get visually overwhelming on large workflows. Error messages are sometimes cryptic without context. And the pricing cliff between Pro and Business on Cloud will frustrate growing teams.

The Sustainable Use License is worth knowing about: it’s not standard open source. Commercial use at scale requires a paid plan. Most individual developers and small teams are well within the free tier, but organizations building n8n-powered products for paying customers should read the license carefully.

Who n8n is built for

n8n fits engineers and technical teams who need workflow automation as part of their stack. Backend developers who want to wire together internal services, send notifications, process webhooks, and run LLM calls without writing a custom service for each task will find n8n a natural fit. DevOps teams building internal tooling around incident response, deployment notifications, or data pipeline automation are a strong match.

It also works well for AI engineers building agent pipelines who want a visual layer that makes debugging easier. If you’ve been writing LangChain scripts and spending too much time on logging and infrastructure, n8n’s AI nodes give you most of the same capability with better observability. For exploring and building AI coding workflows, see the tools listed in best AI agents for coding.

n8n is not built for non-technical users who want to connect apps in five minutes. It’s not a replacement for Zapier if what you need is a quick Slack-to-Google-Sheets sync set up by someone who isn’t a developer. The tool rewards investment.

n8n vs the alternatives

n8n vs Zapier Agents

Zapier is the category default for a reason: it’s polished, fast to start, and works well for common integrations. But it’s designed for non-technical users, and that design shows up as limitations for developers. Zapier doesn’t have code nodes, doesn’t let you self-host, and its per-task pricing gets expensive quickly on high-volume workflows. Zapier’s AI capabilities exist but are shallower than n8n’s full LLM node stack. If you’re technical and need depth, n8n wins. If you need something a non-technical colleague can maintain, Zapier wins.

n8n vs Lindy

Lindy is an AI-first automation tool aimed at business users rather than developers. It’s far easier to get started with and focuses on common business workflows like email triage, meeting summaries, and CRM updates. It doesn’t offer self-hosting, code nodes, or the LLM flexibility that n8n has. The two tools serve different audiences. Lindy is better for a sales or operations team that wants AI assistants without writing a single line of code. n8n is better for a technical team that wants to build those same capabilities with full control over the stack.

n8n vs Gumloop

Gumloop is a newer visual AI workflow builder that leans more into the no-code-friendly direction while still being technically capable. It’s faster to get started with and has a cleaner UI for AI-specific workflows. n8n has a larger integration library, a more mature community, and the self-hosting option that Gumloop lacks. Gumloop makes sense if you want AI workflow automation without infrastructure concerns and don’t need the breadth of integrations n8n offers. n8n makes sense if self-hosting, code nodes, or the existing 500+ integration library are priorities.

Getting started

The fastest path to a running n8n instance is Docker. With Docker installed, one command pulls the image and starts the server on port 5678. From there you create your first workflow in the browser, connect a trigger (webhook, schedule, or manual), add nodes, and run it. The n8n documentation has a Level One course that walks through building a real workflow from scratch, which takes about two hours and covers most of the core concepts.

For Cloud, signup takes a few minutes and the Starter tier includes 50 AI Workflow Builder credits so you can generate a workflow from a text description before you’ve learned the node vocabulary. This is a reasonable on-ramp for teams evaluating n8n before committing to self-hosting.

The community forum and Discord are active. For common questions like connecting a specific API or debugging an authentication error, answers are usually findable in existing threads. The documentation is thorough, though some of the AI node documentation assumes LangChain familiarity.

The bottom line

n8n is what Zapier would be if it were designed for engineers first. It’s genuinely powerful, the AI integration is deep enough to build real agent systems, and the self-hosted Community Edition is free without meaningful strings attached. The learning curve is real and the UI can feel like a mess when workflows get complex, but for technical teams those are manageable trade-offs.

If you’re building automation that needs custom logic, LLM calls, or data privacy guarantees, n8n belongs on your shortlist. If you’re a non-developer wanting to connect two apps quickly, it doesn’t.

Key features

• Native AI nodes: connect OpenAI, Anthropic, Gemini, Mistral, and local models directly in visual workflows
• Code nodes let you write arbitrary JavaScript or Python at any point in a workflow
• Self-host on your own infrastructure with Docker, Kubernetes, or bare metal at no licensing cost
• 500+ pre-built integrations covering Slack, Salesforce, Notion, HubSpot, Google Sheets, and more
• Multiple agent architectures built-in: ReAct, Plan-and-Execute, OpenAI Functions, and Tools Agent
• Visual debugging with per-step input/output inspection and single-step re-runs
• Git-based version control for workflow source management in team environments

Pros and cons

Who is n8n for?

• Developer teams building internal automation that mixes APIs, databases, and LLM calls in one workflow
• Engineers who want to self-host their automation stack for data privacy or cost reasons
• Technical users building AI agents that need human-in-the-loop approvals and full audit visibility
• Ops and DevOps teams automating multi-step processes that require conditional logic and custom code

Alternatives to n8n

If n8n isn't quite the right fit, the closest alternatives are zapier-agents , lindy , and gumloop . See our full n8n alternatives page for side-by-side comparisons.