Building n8n workflows by hand is slow. You hunt for the right node, wire it up, realise a setting is wrong, redo it. The n8n MCP server cuts that loop down to a single prompt.
How it connects
flowchart LR CC[Claude Code] -->|MCP protocol| MCP[n8n MCP Server] MCP -->|reads| DOCS[Node docs\n& templates] MCP -->|calls| API[n8n API] API -->|creates| WF[Deployed workflow]
What the n8n MCP server actually does
The MCP server connects Claude Code directly to your running n8n instance. That connection gives Claude access to the live node documentation, over 3,000 workflow templates, and the n8n API. When you describe a workflow, Claude isn't guessing at node names from memory. It reads the real docs, picks the right nodes, and deploys the workflow straight to your instance.
That's the key difference from pasting a long system prompt into a chat window. System prompts hallucinate node configurations. Direct API access doesn't.
What you need before starting
- A running n8n instance (self-hosted or cloud)
- Claude Code installed
- Node.js
Setting it up
1. Create the MCP config file
In your project root, create a .mcp.json file. Head to the n8n-mcp repository and copy the full configuration block under "full configuration with n8n management tools." You want the management version, not the read-only one, so Claude can actually create workflows.
Fill in two fields: your n8n instance URL (just the base URL, no trailing slash) and your n8n API key. Get the API key from your n8n dashboard under Settings > n8n API.
2. Initialize Claude Code
Open your terminal. Claude Code picks up the .mcp.json automatically. Run /init and say yes to creating the CLAUDE.md file.
3. Add the n8n-specific instructions
From the n8n-mcp repo, copy the "Claude project setup" section and paste it at the end of your CLAUDE.md. Then tell Claude to review the file once. This only needs to happen once per project folder. After that, Claude understands the n8n constraints and won't need reminding.
Building your first workflow
Here's the test prompt from the video:
Use the n8n MCP server to make a workflow. I want a basic chatbot that uses OpenAI, has memory, and can look things up on Wikipedia.
Claude runs, creates the workflow, and prints the URL in your instance. Open it and you'll see the full graph: a chat trigger, an AI agent with a Wikipedia tool attached, a memory module, and the OpenAI model node.
The first run wasn't perfect. The OpenAI credentials weren't connected to the model node. That's expected with AI-generated workflows. Review the output before running it in production.
Use plan mode for better results
Before you hit send on a prompt, press Shift+Tab until "plan mode on" appears. In plan mode, Claude asks clarifying questions before building anything. It catches gaps you didn't notice, like "which OpenAI model do you want?" or "should this trigger on a schedule or on demand?"
Vague prompts produce broken workflows. Plan mode forces you to be specific before any API calls go out. The extra 30 seconds usually saves a 10-minute debugging session.
Who this is for
This tool isn't for beginners who want workflows with zero effort. It's for existing n8n users who know what they're building and want to skip the repetitive node wiring. If you can describe the logic clearly, Claude handles the tedious parts.
What it won't do
It won't catch every misconfiguration on the first try. Generated workflows need a quick review before you run them on live data. Think of it as a fast first draft, not a finished product.
If you're building AI automation workflows for your business and want to skip the trial-and-error phase, get in touch. We build custom n8n workflows for production use.