Back to Blog
GuideJun 30, 202611 min read

n8n MCP Server: Connect n8n to Claude and AI Agents (2026 Guide)

NT

Nikhil Tiwari

MCP Playground

πŸ“– TL;DR

  • n8n speaks MCP two ways: as a server (expose a workflow as tools) and as a client (an AI Agent that calls other MCP servers).
  • The MCP Server Trigger node turns any workflow into an MCP endpoint Claude or Cursor can connect to.
  • The MCP Client Tool node plugs external MCP servers into an n8n AI Agent node as callable tools.
  • You expose tools by wiring normal n8n nodes (HTTP Request, Postgres, Slack…) into the MCP Server Trigger.
  • Test the endpoint against a real model before you connect it to anything in production.

What Is the n8n MCP Server?

n8n is a workflow automation tool that now speaks the Model Context Protocol natively. That means an AI agent like Claude can call your n8n workflows the same way it calls any other MCP tool.

Before this existed, connecting n8n to an LLM meant webhooks and custom glue. You wrote an endpoint, described it by hand, and hoped the model called it right.

MCP removes that friction. n8n exposes a standard, self-describing tool surface, and any MCP client discovers it automatically.

If MCP itself is new to you, read What Is the Model Context Protocol? first β€” this guide assumes you know what a tool call is.

n8n ships two dedicated MCP nodes, and most confusion comes from mixing them up. Let me draw the line clearly.

Two Directions: Server vs Client

n8n can sit on either side of an MCP connection. Which node you reach for depends on who is calling whom.

Node Role Use it when…
MCP Server Triggern8n is the serverYou want Claude/Cursor to call your n8n workflow as a tool
MCP Client Tooln8n is the clientYou want an n8n AI Agent to call an external MCP server

MCP Server Trigger = n8n exposes tools outward. MCP Client Tool = n8n consumes tools from elsewhere. That single distinction unlocks the rest of this guide.

n8n as an MCP Server (MCP Server Trigger)

The MCP Server Trigger node turns a workflow into a live MCP endpoint.

Drop the node onto a canvas and it gives you a URL β€” a Streamable HTTP / SSE endpoint that any MCP client can connect to.

You decide what tools that endpoint exposes by wiring regular n8n nodes into the trigger. An HTTP Request node becomes a tool. A Postgres node becomes a tool. A Slack node becomes a tool.

Each connected node shows up in the client's tools/list response, with its parameters described as a JSON schema the model can reason over.

So you are not writing protocol code. You are dragging the same nodes you already use in n8n, and the trigger handles the MCP plumbing.

Watch the tool descriptions. The model picks tools based on their names and descriptions. A node labelled "HTTP Request1" tells Claude nothing. Rename it to something like "create_invoice" with a clear description.

n8n as an MCP Client (MCP Client Tool)

The MCP Client Tool node does the opposite. It lets an n8n AI Agent node call tools that live on an external MCP server.

Say you built an AI Agent workflow in n8n and you want it to use the GitHub MCP server. You add an MCP Client Tool node, point it at the GitHub MCP endpoint, and the agent can now call those tools.

This is how you give an n8n agent superpowers it does not natively have β€” databases, browsers, third-party APIs β€” without building each integration yourself.

For the orchestration patterns behind multi-tool agents, see multi-agent MCP with CrewAI and LangChain and how MCP tool calling works.

Step by Step: Expose an n8n Workflow as an MCP Server

Here is the shortest path from blank canvas to a working MCP endpoint.

  1. Create a new workflow and add the MCP Server Trigger node.
  2. Copy the MCP URL it generates β€” this is the endpoint clients connect to.
  3. Add a tool node (e.g. HTTP Request or Postgres) and connect it to the trigger's tool output.
  4. Give the tool node a clear name and description β€” this is what the model sees.
  5. Turn on Bearer auth on the trigger and save the token somewhere safe.
  6. Activate the workflow. An inactive workflow will not serve MCP requests.

That last step trips up everyone once. The MCP Server Trigger only responds when the workflow is active β€” a saved-but-inactive workflow returns nothing.

πŸ’‘ Before you paste that URL into Claude, confirm the endpoint actually lists your tools. Test your n8n MCP server free β†’

Connect Your n8n MCP Server to Claude

Once the endpoint is live, connecting Claude is a config edit.

For Claude Desktop, add your n8n endpoint to the MCP config file. Remote HTTP servers go under mcpServers with the URL and your bearer token:

{
  "mcpServers": {
    "n8n": {
      "url": "https://your-n8n-host/mcp/your-workflow-id/sse",
      "headers": {
        "Authorization": "Bearer YOUR_TOKEN"
      }
    }
  }
}

Restart Claude Desktop and your n8n tools appear in the tool list. For the full config-file walkthrough across clients, see the MCP config files guide and how to set up MCP in Claude Desktop.

Using Cursor instead? The flow is nearly identical β€” drop the same URL into Cursor's MCP settings (Cursor MCP setup guide).

Test Your n8n MCP Server Before You Connect It

Problem: you wire up the workflow, paste the URL into Claude, and the model either ignores your tool or calls it with the wrong arguments.

That wastes time, because you are debugging two systems at once β€” n8n and the client β€” with no visibility into either.

The fix is to test the endpoint in isolation first. Open the URL in a browser-based MCP tester, watch tools/list return your tools, and call one by hand.

If the tool names and schemas look right there, the problem is your client config. If they look wrong, the problem is your n8n wiring. You just halved the debugging surface.

Shipping an n8n MCP workflow?

Paste your endpoint into the browser, watch a real model discover and call your tools, and catch schema bugs before Claude ever sees them.

Common n8n MCP Errors & Fixes

The same handful of problems account for most failed connections.

  • Client connects but sees no tools: the workflow is not active, or no tool nodes are wired into the trigger.
  • 401 / Unauthorized: the bearer token in your client config does not match the trigger's token.
  • Model never calls the tool: vague tool names or empty descriptions β€” rename them so the intent is obvious.
  • Tool runs but returns nothing useful: the connected node returns raw data the model cannot parse; shape the output to plain text or clean JSON.
  • Works locally, fails when remote: your n8n instance is not reachable over HTTPS, or a proxy is stripping the SSE connection.

For a broader checklist, see MCP server troubleshooting.

n8n MCP vs Building a Custom MCP Server

n8n is not always the right tool. Pick based on what you are building.

Use the n8n MCP Server when…

  • The logic is already a workflow β€” approvals, data syncs, notifications.
  • You want non-developers to edit what the agent can do.
  • You need a dozen integrations stitched together fast, with no SDK code.

Build a custom MCP server when…

  • You need fine control over tool schemas, errors, and streaming.
  • The server is a product surface, not an internal automation.
  • Performance and versioning matter more than drag-and-drop speed.

If you land on the custom route, start with the Node.js MCP server guide or the Python FastMCP tutorial. You can also wrap your existing API β€” see wrapping APIs as MCP tools.

Already automate with Zapier? The same idea applies there β€” compare with Zapier MCP for Claude.

Frequently Asked Questions

What is the difference between the MCP Server Trigger and MCP Client Tool in n8n?+
The MCP Server Trigger makes n8n a server: it exposes your workflow's nodes as tools that an external client like Claude can call. The MCP Client Tool does the reverse: it lets an n8n AI Agent node call tools hosted on an external MCP server. One serves tools out, the other consumes tools in.
Can Claude call an n8n workflow directly?+
Yes. Add an MCP Server Trigger to the workflow, activate it, and add its URL and bearer token to Claude's MCP config. Claude then discovers the workflow's tools and can call them like any other MCP tool.
Why does my n8n MCP server return no tools?+
The most common cause is an inactive workflow β€” the MCP Server Trigger only responds when the workflow is active. The second most common cause is having no tool nodes wired into the trigger's tool output. Activate the workflow and connect at least one tool node.
Do I need to self-host n8n to use MCP?+
No. Both the self-hosted and cloud versions of n8n support the MCP nodes. Self-hosting gives you more control over networking and the endpoint URL, which can matter when exposing the server to remote clients over HTTPS.
How do I test an n8n MCP server before connecting it to Claude?+
Paste the MCP Server Trigger URL (with its bearer token) into a browser-based MCP tester. Confirm tools/list returns your tools with sensible names and schemas, then call one manually. If that works, any remaining issue is in your client config rather than n8n.

Related Guides

Further Reading

πŸ€– Test your n8n workflow against a live AI model

Before Claude depends on your workflow, make sure the tools actually fire. Use MCP Agent Studio to chat with your n8n MCP server and watch a real model call its tools. Works with Claude, GPT, Gemini and more. Free credits on sign-up.

NT

Written by Nikhil Tiwari

15+ years in product development. AI enthusiast building developer tools that make complex technologies accessible to everyone.

Build, compare & ship MCP agents β€” free

Connect any MCP server, compare 40+ models side-by-side, deploy hosted servers, and save reusable agents you can export as an API β€” all in your browser.

✦ Free credits on sign-up · no credit card needed

Try for Free β†’
n8n MCP Server: Connect n8n to Claude and AI Agents (2026 Guide)