Drive n8n workflows with Claude MCP — from setup to real use

If your day is open the n8n canvas, drag a node, fill parameters, delete the node, repeat — Claude MCP takes that click cycle off your hands. Just say, in English, "build me a workflow that posts a sales summary to a Slack channel."

MCP (Model Context Protocol) is an open protocol Anthropic released in late 2024, defining how LLMs talk safely to external tools and data sources. The Claude MCP ecosystem exploded in the first half of 2026, and driving automation tools like n8n through conversation is now stable enough for real work.

Why Claude MCP × n8n, and why now

Three things aligned at once. First, the MCP spec stabilized at 1.0 in early 2026, locking down the three axes of tool calls, resources, and prompts. Second, n8n 2.29+ was refactored to be API-first, exposing workflow CRUD, execution triggers, and node schemas over REST. Third, the community-built @czlonkowski/n8n-mcp server matured to production grade, so you can wire it up without writing anything yourself.

What you need

• An n8n 2.29+ instance (self-hosted or Cloud, either works)
• An n8n API key — issued from Settings → API, scope at minimum workflows and executions
• Latest Claude Desktop (or the Claude Code CLI)
• Node.js 20+ — the MCP server runs locally on demand
• Other MCP clients (Cursor, Continue, etc.) work with the same spec

30-minute setup

1. In n8n, go to Settings → API, click Create API key, pick scope workflows,executions,credentials:read, and save the key immediately (it's shown only once)
2. In your terminal, run npx @czlonkowski/n8n-mcp install — this installs the server script locally
3. Open the Claude Desktop config file claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/)
4. Add the mcpServers block below and swap YOUR_N8N_URL and YOUR_N8N_API_KEY for real values
5. Fully quit and restart Claude Desktop (⌘Q, then relaunch — reloading the window is not enough)
6. Success looks like a green dot next to n8n-mcp under Settings → Developer
7. Type show me my n8n workflows in the chat to confirm the first tool call
8. If that works, ask it in natural language to build any workflow

{
  "mcpServers": {
    "n8n-mcp": {
      "command": "npx",
      "args": ["-y", "@czlonkowski/n8n-mcp"],
      "env": {
        "N8N_URL": "YOUR_N8N_URL",
        "N8N_API_KEY": "YOUR_N8N_API_KEY"
      }
    }
  }
}

First conversation — ask it to build a workflow

Once setup is done, it's just a conversation. Write what you want in natural English. Here's a prompt I actually pasted.

Build a workflow that posts yesterday's sales summary to the Slack #alerts channel every morning at 9am.

Claude moves in three steps. First it scans available nodes with the list_nodes tool and picks a Schedule Trigger + Postgres + Slack combination. Then it assembles JSON and hands it off to the create_workflow tool. Finally it reports the saved workflow ID and edit URL back into the chat. The whole thing wraps in about 90 seconds.

'Build a thing that posts a sales summary to Slack' replaced 40 clicks. I was skeptical at first, but the workflow it made just runs.— Kim Jihoon · Marketing

Real usage patterns

• New workflow drafts — describe the requirement in a sentence and a draft appears on the canvas. Humans only review and tune parameters
• Refactor existing workflows — handle requests like switch this workflow's failure alerts from Slack to Telegram
• Summarize execution logs — natural-language write-ups of last night's failures, pulling logs via the executions tool
• Bulk node parameter edits — update the Postgres host across many workflows in one shot
• Error investigation — ask why a specific workflow keeps failing and it walks through recent runs and per-node errors
• Workflow documentation — for handoff, ask it to render a workflow's structure as markdown

FAQ

• Q. Does it work with Cloud instances? A. Yes. As long as you have an API key, self-hosted and n8n Cloud behave identically. Only the URL changes to https://your-org.app.n8n.cloud.
• Q. Where does the MCP server run? A. Claude Desktop launches it locally via npx automatically. No separate server deployment needed. If you want to move it to the cloud, swap the stdio transport for sse.
• Q. Can I scope permissions narrowly? A. Yes, via n8n API key scopes. For example, read-only lets it query and summarize but blocks create/delete. Recommended for production.
• Q. What if I accidentally delete a production workflow? A. @czlonkowski/n8n-mcp puts a confirmation prompt on delete tools, but the real answer is automated backups. Build one dedicated backup workflow.
• Q. Can I converse in Korean? A. Perfectly. Claude calls tools regardless of prompt language and returns summaries in Korean.
• Q. Do other MCP clients work? A. Claude Code CLI, Cursor, Continue — any client that follows the MCP spec plugs in the same way. Only the config file location differs.
• Q. What about Zapier or Make? A. Only the n8n community server is at production maturity right now. MCP servers for Zapier and Make are early — not ready for real work yet.

Limits and caveats

• Human review is still required — even a workflow Claude built needs a human to hit the activate button, and you should eyeball it on the canvas at least once before it touches production data
• API rate limits — the n8n API caps requests per minute, so batch large bulk operations
• Split write-only keys in production — separate editing sessions from automated runs with different API keys to limit blast radius
• The MCP server itself doesn't handle user Q&A — answering user questions during a workflow run is the job of a separate AI Agent node. See: n8n in 2026: AI Agent nodes are now the workflow default

The Claude MCP × n8n combination is a preview of the next stage of workflow automation. From placing nodes by hand to describing the outcome in a sentence. New to n8n? Start with Webhooks 101. Already running it? Spend 30 minutes today and wire up MCP.