n8n Review 2026: Open-Source Zapier With AI Agents and MCP

n8n (pronounced “n-eight-n”) is the automation platform we reach for when a Zapier bill starts looking unreasonable, when a workflow needs real code, or when an AI agent has to do more than send an email. It’s source-available, deployable anywhere — Cloud, Docker, npm, desktop — and in 2026 it has quietly become the most credible AI-agent automation layer outside the big-cloud ecosystems. This is the honest review: where n8n is genuinely brilliant, where it punishes the unprepared, and how to decide whether it belongs in your stack.

What makes n8n different from Zapier in 2026

The shortest version: Zapier is a SaaS product. n8n is a piece of software you can run anywhere. That single architectural choice cascades into almost every other difference between the two tools — pricing, flexibility, AI capability, and learning curve.

Zapier is the right answer for a non-technical operator who needs to wire ten apps together and never wants to think about infrastructure. It has 7,000+ pre-built integrations and a setup flow that someone in marketing can follow on a Tuesday afternoon. We respect that, and we cover it in our Zapier guide.

n8n is built on a different bet: that the most valuable automations in 2026 involve real logic, real data transformations, and real AI agents — and that those automations deserve a tool with code nodes, sub-workflows, and a self-hosted option. The n8n canvas looks like Zapier’s at a glance, but the moment you need a loop, a branch, a Python expression, or an LLM that calls back into the workflow, the difference becomes obvious.

Three concrete distinctions matter most:

• Pricing model. n8n charges per execution — one full workflow run, regardless of how many nodes it touches. Zapier charges per task — every action step counts. A 20-node n8n workflow processing 500 records is one execution. The same logic in Zapier could be hundreds of tasks. On heavy workloads the cost gap is roughly 90%.
• Self-hosting. n8n’s Community Edition is free, open under the Sustainable Use License, and runs on your own Docker host with unlimited executions. Zapier has no equivalent — you’re a SaaS customer or you’re nothing.
• Code and AI depth. Native JavaScript and Python code nodes, mature LangChain integration, six different agent node types, and bidirectional MCP support. Zapier’s agent feature is newer and intentionally simpler.

Honest counterweight: n8n has roughly 450-500 integrations versus Zapier’s 7,000+. If you need a niche SaaS connector — some specific CRM nobody outside its category has heard of — Zapier will probably have it and n8n probably won’t. You can fall back to a generic HTTP node, but that’s real work.

Workflow building: nodes, code nodes, sub-workflows, error workflows

An n8n workflow is a directed graph of nodes connected on a visual canvas. A trigger fires (webhook, schedule, manual, form, app event), data flows from node to node, and each node does one thing — pull a record, transform a field, call an API, write to a database. That part is familiar to anyone who has used a no-code tool before.

What sets n8n apart is what happens when the visual canvas isn’t enough:

• Code nodes. Drop in a JavaScript or Python node and write whatever you want. Run the code once per item or once per workflow. This is the escape hatch you reach for when an expression isn’t enough — JSON wrangling, regex, custom auth headers, weird date math, or calling an unsupported library.
• Expressions. Inline templating with $json, $node, and $workflow references. Powerful, but be warned: the syntax has a learning curve. A beginner will spend a real afternoon getting comfortable with how data shape changes between nodes.
• Sub-workflows. Call one workflow from another. This is how serious n8n users keep things sane — extract reusable pieces (auth, logging, error handling) and call them like functions.
• Error workflows. A separate workflow that fires whenever a parent workflow fails. Lets you Slack the team, write to a log, retry, or open a ticket. This is one of the features that quietly makes n8n production-grade.
• Branching, merging, loops. IF, Switch, Merge, SplitInBatches, and Loop nodes give you full control flow. Combined with sub-workflows, you can build automations that genuinely deserve to be called software.
• Human-in-the-loop. Wait nodes and form triggers let you pause a workflow for approval — useful for AI agents that should not be allowed to send the email themselves.

The debugging tooling is a real strength. Single-step re-runs, data replay through any node, and full execution logs let you inspect exactly what data hit a node and what came out. When a Zapier zap silently breaks, you stare at a vague failure message; when an n8n workflow breaks, you click the failed node and see the exact payload.

AI nodes: LangChain agents, vector stores, RAG, evals

This is where n8n has pulled meaningfully ahead of every other no-code automation tool. The AI / LangChain cluster nodes turn the canvas into a serious agent-building environment without forcing you into Python.

The cluster covers the full stack:

• Language models. OpenAI, Anthropic Claude, Google Gemini, Ollama, and any local model that exposes a compatible endpoint. Swap providers by changing one node.
• Embeddings and vector stores. Pinecone, Qdrant, Supabase, and PGVector are first-class. Pair an embedding node with a vector store node and you have RAG in about ten minutes.
• Retrievers, output parsers, memory. The pieces you need to make an agent useful across multiple turns of a conversation.
• Agent nodes. Six variants — Tools Agent (the new default), OpenAI Functions Agent, Plan-and-Execute Agent, ReAct Agent, SQL Agent, and Conversational Agent. Pick the one that matches the shape of your problem.
• Evaluations. Built-in evals tooling lets you grade agent outputs against test cases. This is the feature that separates “I built a demo” from “I shipped this to production.”

The reason this stack matters: n8n agents can call any other n8n node as a tool. That means an agent can hit your CRM, query your database, post to Slack, write to a Google Doc, and trigger a sub-workflow — all from a single prompt. Pair that with sensible prompt design and you get agents that do real work, not just chat.

Honest note: agent reliability is still hard. Tools Agent is the most stable of the six, but you will spend time tuning system prompts, choosing the right model (we lean Sonnet or GPT-4-class for the brain, Haiku-class for the cheap tool calls), and writing evals. n8n makes the plumbing easy. It does not make the agent smart.

MCP integration: bidirectional client + server (the 2026 standout)

If you only remember one thing from this review, remember this section. n8n is the first mainstream automation platform to ship full bidirectional Model Context Protocol support, and that capability is genuinely ahead of the field in 2026.

MCP is the open protocol — backed by Anthropic, adopted across the major coding agents — that lets AI clients discover and call external tools in a standard way. n8n implements both sides:

• MCP Client Tool. Point an n8n agent at any MCP server URL and it auto-discovers and invokes every tool that server exposes. You inherit a whole ecosystem of MCP servers (file systems, databases, GitHub, Slack, Linear, hundreds more) without writing integration code.
• MCP Server Trigger. Turn any n8n workflow into an MCP tool. Claude Desktop, Cursor, VS Code, and any other MCP-aware client can now call your workflow as if it were a native tool. This is how you give a coding agent the ability to “ping our staging environment and roll back if the smoke tests fail” without writing a custom integration.
• n8n’s own MCP server. n8n itself ships an MCP server that lets coding agents author, validate, test, and deploy n8n workflows from a single prompt. Ask Claude in your editor to “build me a workflow that watches a Gmail label and summarizes new threads to a Slack channel” and it can actually do it.

The strategic implication is bigger than it looks. Most automation tools assume the human is the one orchestrating. n8n’s MCP stack assumes the human or an AI agent can be the one orchestrating, on either side of the wire. That bet looks more correct every month.

Self-hosted vs Cloud: when each makes sense

Both deployment paths are first-class. The right choice depends on volume, ops capacity, and how much your data hates leaving your perimeter.

Choose Cloud when:

• You’re under ~10,000 executions per month and the math on Pro tier works.
• You don’t have someone on the team who is comfortable with Docker, Postgres, and Redis.
• You want managed updates, backups, EU/US data regions, and SOC-style compliance without doing it yourself.
• You’re prototyping and want zero setup friction.

Choose self-hosted Community Edition when:

• Your execution volume is high enough that Cloud pricing hurts. The crossover is real around 40-50K executions/month.
• Data residency or privacy rules say workflows can’t run on a third-party host.
• You already have a Docker host and someone who knows what a queue worker is.
• You want unlimited executions for free under the Sustainable Use License.

The honest warning on self-hosting: this is real ops work. Docker, Postgres, Redis (for queue mode), backups, TLS certificates, version upgrades, and monitoring are all your problem. n8n’s docs and Helm chart make it tractable, but if you can’t tell us what your Postgres backup strategy is, stay on Cloud until you can.

A middle option exists: Enterprise self-hosted licensing adds SSO, RBAC, log streaming, external secrets, audit logs, and the right to embed n8n inside your own product. For mid-size teams that need governance, this is worth the conversation.

Pricing: Cloud tiers vs self-hosted Community Edition

n8n’s Cloud pricing is in EUR on the official site. The 2026 tiers, on annual billing:

• Starter — EUR 20/mo (~$24). 2,500 executions/mo, 5 concurrent runs, 1 shared project. The right tier for hobbyists, side projects, or proving the platform.
• Pro — EUR 50/mo (~$60). 10,000 executions/mo, 20 concurrent, 3 shared projects, admin roles, global variables, 7-day insights. The natural home for most small teams.
• Business — EUR 667/mo. 40,000 executions/mo, 30 concurrent, 6 shared projects, SSO/SAML/LDAP, 30-day insights, Git version control, self-hosted option. Genuine enterprise readiness.
• Enterprise — custom. Custom volume, 200+ concurrent, unlimited shared projects, 365-day insights, dedicated SLA, external secret store.

Self-hosted:

• Community Edition — free. Unlimited executions under the Sustainable Use License. Permits internal business use; restricts reselling the software as a service. For most teams, this is the most cost-effective option that exists in the category.
• Enterprise self-hosted — paid. Adds SSO, RBAC, log streaming, external secrets, audit logs, and commercial-use rights for embedding inside other products.

The honest critique: the jump from Pro (EUR 50) to Business (EUR 667) is brutal. If your team needs SSO, Git version control, or more than 10K executions but doesn’t need 40K, you’re stuck overpaying or self-hosting. n8n knows this is awkward; it has not yet been fixed. For mid-size teams, the right move is often “Pro on Cloud while you grow, then jump to self-hosted Community Edition or Enterprise self-hosted when you outgrow it” — skipping Business entirely.

The execution-vs-task pricing model genuinely matters. A workflow that hits 20 nodes to process one record is one execution on n8n. The same shape on Zapier could be 20 tasks. At scale, this is why operators with serious volume migrate.

n8n vs Zapier vs Make vs Pipedream

Four tools, four different bets about who the customer is. Quick read on each:

• Zapier. Widest integration coverage (7,000+), easiest setup, best for non-technical users. Per-task pricing makes it expensive at scale. Agent feature is real but newer and simpler than n8n’s. Our full take is in the Zapier guide.
• Make (formerly Integromat). Strong visual scenario builder with detailed control over data flow. Sits between Zapier and n8n on the technical-comfort axis. No self-hosting. Pricing is per-operation, similar feel to per-task.
• Pipedream. Code-first by design. JavaScript and Python steps as the primary unit. Excellent for developers who want a serverless-feeling automation tool. Smaller integration library than Zapier; weaker visual canvas than n8n or Make.
• n8n. The most flexible of the four. Self-hosting, code nodes, the deepest AI agent stack, MCP support, execution-based pricing. Steeper learning curve. Smaller integration library than Zapier.

A simple decision framework: If your team is non-technical and you need 7,000 integrations, use Zapier. If your team is comfortable with code and you want maximum flexibility plus AI-agent depth, use n8n. If you want a code-first developer experience without the visual canvas, look at Pipedream. Make is the strong middle option if Zapier is too expensive but you’re not ready for n8n.

For more on where automation fits in a broader AI stack, see our AI tools directory.

Getting started: your first agent workflow in 30 minutes

Skip the tutorial about Slack-to-Google-Sheets. Build something interesting on day one — an AI agent that summarizes new emails into a daily digest. Here’s the path that works:

1. Sign up for Cloud Starter at n8n.cloud (or run docker run -p 5678:5678 n8nio/n8n if you want self-hosted). The cloud option saves you setup time on day one.
2. Create a new workflow. Add a Schedule Trigger node, set it to run once a day at 7am.
3. Add a Gmail node. Configure OAuth to your account, set the operation to “Get Many” with a label or date filter so you only pull yesterday’s important threads.
4. Add an AI Agent node (Tools Agent). Connect a Chat Model — Claude Sonnet 4.6 or GPT-4-class. Write a system prompt that says “You are a daily-digest assistant. Read the messages in the input. Output a markdown summary grouped by sender, highlighting anything that needs a reply today.”
5. Add a Send Email or Slack node that delivers the summary to you.
6. Test it. Click Execute Workflow. Open each node in the canvas and inspect the data shape — this is where you’ll learn how $json works.
7. Add an error workflow. Settings → Error Workflow, point it to a tiny second workflow that pings you on Slack if the digest fails.
8. Activate it. Toggle the workflow on. Tomorrow morning, you have a working AI agent.

From there, the natural next steps are: add a vector store node so the agent has memory, expose the workflow as an MCP tool so Claude Desktop can trigger it on demand, or chain a sub-workflow that drafts replies for the threads that need them.

The honest takeaway. n8n is more powerful than Zapier and steeper than Zapier. Both halves of that sentence are true and neither is going to change. If you’re a non-technical operator who just needs three reliable zaps, stay on Zapier and don’t feel bad about it. If you’re building AI agents, running real volume, or you can spell “Docker” without flinching, n8n is one of the most leveraged tools you can put in your stack in 2026 — and the Community Edition makes the price of finding out roughly zero.