Automation & Workflows · Blog Post
n8n Skills for Claude Code: building n8n workflows through n8n-mcp
Giving Claude access to n8n through an MCP server is the easy part. Getting it to configure nodes correctly, write expressions that don't silently resolve to null, and stop looping on validation errors is the hard part. This pack, from the author of n8n-mcp, is 14 skills plus a router and a set of hooks aimed squarely at that gap.
By Mehadi Hasan 8 min read
What it is
n8n-skills is a skill pack by Romuald Członkowski, who also maintains n8n-mcp, the open-source MCP server that lets AI assistants search n8n nodes, create and edit workflows, validate them, read failed executions, and manage credentials. The skills are the other half of that setup. The server gives Claude the tools. The skills teach it how to use them without the usual mistakes.
The README lists the problems it targets plainly: using MCP tools incorrectly, getting stuck in validation error loops, not knowing which workflow pattern to use, and misconfiguring nodes and their dependencies. Anyone who has watched an agent build an n8n workflow will recognize all four. The classic one is webhook data: it lives under $json.body, not $json, and two separate skills flag it as a critical gotcha.
Older write-ups describe it as a 7-skill pack. The current repo (plugin version 1.37.1, last pushed October 24, 2026) has grown to 14 capability skills, an always-on router skill, and a hooks enforcement layer. It is MIT-licensed. The hook scripts are adapted from n8n's own official n8n-io/skills project under Apache-2.0, credited in the repo's NOTICES file.
What's inside
The 14 skills group naturally into five areas. Each is a SKILL.md with supporting reference files (error catalogs, pattern libraries, a native-node guide), and the repo ships evaluation scenarios for every skill.
- Using the MCP server (n8n-mcp-tools-expert, n8n-validation-expert, n8n-multi-instance): Which n8n-mcp tool to call for which job, the nodeType format differences, validation profiles, which warnings are false positives, and how to name the target instance when one account reaches several n8n instances.
- Designing workflows (n8n-workflow-patterns, n8n-node-configuration, n8n-subworkflows, n8n-error-handling): Five architectural patterns (webhook, HTTP API, database, AI, scheduled), operation-aware node setup, typed sub-workflow inputs, and error outputs, retries, and 4xx/5xx response shapes.
- Data and code (n8n-expression-syntax, n8n-code-javascript, n8n-code-python, n8n-code-tool, n8n-binary-and-data): The {{ }} expression rules, Code node data access and return formats, native Python sandbox limits, the AI-agent Custom Code Tool (a different node with a different contract), and $binary vs $json.
- AI agents (n8n-agents): Agent vs LLM chain vs Text Classifier, the model/memory/tools/output-parser slots, $fromAI, structured output, RAG, and human-in-the-loop review.
- Deployment (n8n-self-hosting): Deploys self-hosted n8n to a fresh Linux VM over SSH with Docker Compose behind Caddy, in single or queue mode. It sits outside the router and hooks flow.
The fifteenth folder, using-n8n-mcp-skills, is the router. It doesn't hold reference material. It tells Claude which specialist skill owns the decision it is about to make, summarizes every n8n-mcp tool, and states three non-negotiable rules: invoke the relevant skill before any n8n action, validate and then inspect the saved workflow before activating, and never put secrets in text fields.
How it works
Skills load on their descriptions, as with any SKILL.md pack. Ask "build a webhook to Slack workflow" and the README walks through how several compose: workflow-patterns picks the webhook pattern, mcp-tools-expert finds the nodes, node-configuration sets them up, expression-syntax handles the mapping, and validation-expert reads the results.
What sets this pack apart is the hooks layer, which fires at the moment of decision instead of waiting for a prompt to match a description:
- SessionStart: Loads the using-n8n-mcp-skills router into every session, and again on resume, clear, and compact so it survives context compaction
- PreToolUse: Before get_node, create, update, validate, test, n8n_instances, and n8n_manage_credentials calls, prints a short reminder pointing at the skill that owns that decision
- PostToolUse: After validate_workflow, reads the node types in the workflow and routes Claude to the skills that cover the remaining risks
Some details are well thought out. Lookups of high-risk nodes (Set, Code, Merge, Loop Over Items, Date & Time, Data Table, AI Agent) fire their reminder every time, not once per session, because a repeat lookup usually means another instance of the same decision. Create and update calls are scanned for Code or HTTP Request nodes missing a Why Code: or Why HTTP: note. Every hook fails open and never blocks a tool call. They are bash scripts that need jq, and exit quietly without it.
The guidance has a consistent opinion running through it: native node first, Code last. The router has a job-to-node table (Summarize for group-and-sum, Remove Duplicates for dedupe, Compare Datasets for list diffs, HTTP Request's built-in pagination for paging) and a closed list of cases where a Code node is justified.
The router's transform order
1. the service's own node (HTTP Request only if none covers it) 2. a native node for the job (Summarize, Merge, Sort, Crypto ...) 3. a short expression inside that node 4. a Code node, with a "Why Code:" note explaining why
The prerequisite: n8n-mcp
This is the part to get right first. The installation guide is blunt: you must have the n8n-mcp server installed and configured before using these skills. The skills are written around its tool names and parameter formats, so without it they are mostly reading material.
You have two options. The n8n-mcp README offers a hosted service at dashboard.n8n-mcp.com, run by the same author, where you connect your n8n instance and point your MCP client at it. Or you self-host it with npm, Docker, or Railway. For a local setup in Claude Code, the repo's example config looks like this (placeholders, not real values):
.mcp.json: self-hosted n8n-mcp
{
"mcpServers": {
"n8n-mcp": {
"command": "npx",
"args": ["n8n-mcp"],
"env": {
"MCP_MODE": "stdio",
"LOG_LEVEL": "error",
"DISABLE_CONSOLE_OUTPUT": "true",
"N8N_API_URL": "https://your-n8n-instance.com",
"N8N_API_KEY": "your-api-key-here"
}
}
}
}
The API URL and key are optional. Without them Claude can search nodes and templates and validate configurations. With them it can create, update, and test workflows on your instance, which is what most people want. Your n8n instance needs the Public API enabled. The README recommends n8n-mcp 2.97.0 or later, since that release added what the skills' native-node guidance builds on. Older servers still work.
Worth knowing: n8n now ships its own built-in MCP server. The n8n-mcp README is candid about the trade-off. The built-in one is enough for a single instance on a recent n8n 2.x where you set up credentials yourself. n8n-mcp adds multi-instance support, credential management, autofix, execution error analysis, templates, and a security audit.
Installation
The README's recommended route is the Claude Code plugin, which is the only install that brings the hooks:
Claude Code: plugin install
/plugin install czlonkowski/n8n-skills
Or add the repo as a marketplace, then pick n8n-mcp-skills from the list:
Claude Code: via marketplace
/plugin marketplace add czlonkowski/n8n-skills /plugin install
The manual route copies the skill folders into your skills directory. You get all 15 skills, but no hooks:
Manual install (macOS / Linux)
git clone https://github.com/czlonkowski/n8n-skills.git cp -r n8n-skills/skills/* ~/.claude/skills/
Outside Claude Code, the README covers uploading zipped skill folders to Claude.ai (Settings → Capabilities → Skills, with prebuilt zips on the latest release), and a release bundle of .skill files for ChatGPT and other hosts that read SKILL.md. The hooks also run in the Codex plugin install. On every other host the skills still trigger on their descriptions, just without the nudges.
Using it day to day
There are no commands to learn. You ask for n8n work in plain language ("find me a Slack node", "why is validation failing?", "how do I access webhook data in a Code node?") and the matching skill loads. With the plugin installed, the router is already in context from the first turn.
The rule that matters most in practice is the router's second one: validate and verify. After every create or update, Claude is told to fetch the workflow back and inspect its connections object, because validation alone misses dropped wires, Merge index mix-ups, and error outputs that were never connected. The skills also warn that a green test run isn't proof. A JavaScript error inside {{ }} resolves to null, so a Filter with a broken condition drops every item and still shows success. That is the gap between "the JSON is valid" and "the workflow works", which I cover in more depth in what makes an n8n workflow production-ready.
Two skills cover jobs beyond building. n8n-multi-instance matters if one n8n-mcp account reaches production, staging, or several client instances: it teaches Claude to name the target instance on every call, especially credential writes. n8n-self-hosting deploys n8n to a fresh Linux VM over SSH. It asks single or queue mode first, then generates secrets on the server.
Pricing
The skills are free and MIT-licensed. n8n-mcp is also MIT-licensed and free to self-host. The hosted n8n-mcp service has a free tier of 100 tool calls a day, according to its README. An agent building and validating a workflow makes a lot of tool calls, so heavy users will likely outgrow that and either self-host or move to a paid plan.
The other costs are the ones you already have: your Claude plan or API usage, and wherever your n8n runs. If you are weighing n8n Cloud against self-hosting, my breakdown of what n8n automation actually costs covers the numbers.
My take
If you already drive n8n from Claude through n8n-mcp, this pack is an easy add. It is written by the person who wrote the tools it teaches, and it shows: the guidance is built around real parameter formats, real error strings, and specific n8n traps (the Code Tool vs Code node return formats, responseCode defaulting to 200, Merge dropping binary data) rather than generic advice. The native-node-first stance is the right default: a Summarize or Merge node is easier for the next person to read and debug in the n8n editor than a Code node doing the same job.
The trade-offs are real, though. It is a large pack, and the router tells Claude to load more skills rather than fewer, which costs context on every n8n task. The README's own headline numbers (525+ nodes, 2,653+ templates) lag behind the current n8n-mcp README, a small sign of how fast this area moves. The router admits as much and tells Claude to trust the live tool over the skill when they disagree. And skills reduce mistakes; they don't remove them. An agent with write access to your n8n instance and credential tools still needs review before anything gets activated, ideally on a staging instance first.
Who it suits: people comfortable with Claude Code and MCP config who build n8n workflows regularly and want fewer correction loops. If you'd rather have production workflows designed, built, and monitored for you, that is what my n8n automation service does.
Frequently asked questions
Do I need the n8n-mcp server to use these skills?
Yes. The skills teach Claude how to use n8n-mcp's tools (search_nodes, get_node, validate_workflow, n8n_update_partial_workflow and the rest), so without the server they are mostly reference text. You can self-host n8n-mcp for free via npm or Docker, or use the author's hosted service, which has a free tier of 100 tool calls a day. Add your n8n API URL and key if you want Claude to create and edit workflows, not just look up nodes.
What is the difference between the plugin install and copying the skill folders?
The plugin install brings the hooks layer: a SessionStart hook that loads the router skill every session, PreToolUse reminders before high-impact n8n-mcp calls, and a PostToolUse check after validate_workflow. Copying the folders into ~/.claude/skills gives you the 15 skills without the hooks, so they only activate when a request matches a skill description.
Does it work on n8n Cloud or only self-hosted n8n?
The workflow-building skills work with any n8n instance that n8n-mcp can reach through the n8n Public API, Cloud included. The one exception is the n8n-self-hosting skill, which deploys n8n to your own Linux server with Docker Compose and Caddy and is explicitly not for n8n Cloud. The Python skill also notes that imports are blocked on n8n Cloud.
Can I use it outside Claude Code?
Partly. The README covers uploading the skills to Claude.ai, and a release zip of .skill files for ChatGPT and other hosts that read SKILL.md. The hooks only run in the Claude Code or Codex plugin install. Everywhere else the skills still trigger on their descriptions, just without the proactive reminders.