Connect n8n to Claude and ChatGPT with MCP

#01Why connect n8n to an assistant?

Marketers already ask Claude and ChatGPT questions all day. If those assistants can call your n8n workflows, they can answer with your real data: last week's KPIs, current AI visibility, the latest competitor changes. You control exactly what each tool does, because the logic lives in your workflow.

The Claude MCP connectors lesson explains MCP from the assistant side. This lesson covers the n8n side.

#02The two directions

#03Setting up the MCP Server Trigger

According to the n8n docs, the MCP Server Trigger has a test URL and a production URL. Test mode shows calls in the editor, while production runs when the workflow is published. The trigger only connects to tool nodes, not to regular nodes. To expose your own workflow logic, use the Custom n8n Workflow Tool.

Authentication options include Bearer, Header auth and, from version 2, n8n User Auth via OAuth2. The docs also describe how to connect Claude Desktop through a gateway (they show mcp-remote) for clients that need it.

If you self-host behind a reverse proxy, the docs note SSE needs proxy buffering turned off.

#04Using the assistant side

In Claude, you add custom connectors with a remote MCP server URL; available options depend on your plan, so check Anthropic's current help pages. ChatGPT supports custom MCP connectors on some plans. The ChatGPT track's work agent lesson covers ChatGPT-side workflows.

Once connected, ask questions in plain language. The assistant reads tool descriptions to decide when to call them, so clear names and descriptions matter.

Our view: MCP is the easiest way to give your team AI access to clean marketing data without giving the AI direct access to every tool. Keep the logic and permissions in n8n.

#05Safety rules

• Use Bearer or OAuth, never None, for real data.
• Read-only tools first. Write tools need human review (lesson 9) or shouldn't be exposed.
• Don't return personal data unless the use is approved.
• Rotate tokens when people leave.

#06MCP Server Trigger settings

#07Tools that suit MCP for marketers

Return compact JSON. Long outputs cost tokens and confuse the assistant.

#08Writing tool descriptions

Assistants decide when to call a tool from its name and description. Good descriptions say:

• What it returns: "Last full week's sessions, key events and paid spend for Ledgerline, Monday to Sunday, Berlin time."
• When to use it: "Use for questions about recent marketing performance."
• What it doesn't cover: "No data older than 12 weeks; no revenue."
• Inputs and formats: "Optional week_start as YYYY-MM-DD."

#09Troubleshooting connections

Start with one tool, then add more as people ask for them.

#10A sub-workflow tool, step by step

The tool the assistant calls is a normal workflow:

In the MCP workflow, the Custom n8n Workflow Tool points at this workflow. Give it a name such as get_weekly_kpis and a description saying what it returns and the accepted input. Use the tool's input schema so the assistant knows week_start is optional and formatted YYYY-MM-DD.

#11Client configuration

For clients that only support local stdio servers, the n8n MCP docs show a gateway setup using mcp-remote in the Claude Desktop config, along these lines:

{
  "mcpServers": {
    "n8n": {
      "command": "npx",
      "args": ["mcp-remote", "YOUR_PRODUCTION_URL", "--header", "Authorization: Bearer ${AUTH_TOKEN}"],
      "env": { "AUTH_TOKEN": "your-token" }
    }
  }
}

Check the docs for the exact current format. Clients that support remote MCP servers directly can use the production URL without a gateway.

#12Cost math

Each tool call is a workflow execution: the MCP Server Trigger workflow and the sub-workflow it calls. If ten people each ask three questions a day on working days, that's around 600 tool calls a month. On Starter's 2,500 monthly executions, that's a meaningful share, so check how your plan counts sub-workflow runs and watch usage in the first month.

#13Ledgerline walkthrough: rollout

Ledgerline is a fictional accounts-payable software company. Details are illustrative.

Marketing ops piloted the tools with three people for two weeks before sharing the URL more widely. They rotated the Bearer token once after a pilot user pasted their config into a shared document by mistake. The log table showed most questions were about last week's numbers, so they added a 30-minute cache in a data table to avoid repeated GA4 calls.

#14Worked example: Ledgerline's KPIs inside Claude

Ledgerline is a fictional accounts-payable software company. Details are illustrative.

Ledgerline's team used Claude daily. Marketing ops exposed three read-only tools through an MCP Server Trigger with Bearer auth: weekly KPIs, AI visibility and competitor changes.

The first version returned whole GA4 reports, which made answers slow and rambling. Trimming outputs to a few fields fixed it. A log table showed the KPI tool used most, mainly before Monday stand-ups. A request for a "create campaign" tool was declined until a human review step could be added.

Walkthrough

1. Build a sub-workflow "get_weekly_kpis" that starts with When Executed by Another Workflow and returns a small JSON object.
2. Create a new workflow with an MCP Server Trigger. Set authentication to Bearer and generate a long token.
3. Attach a Custom n8n Workflow Tool pointing to get_weekly_kpis, with a clear name and description.
4. Write Instructions: "Tools for Ledgerline marketing data. Numbers are weekly, Berlin time."
5. Test with the test URL from an MCP client. Then publish and use the production URL.
6. Connect from Claude using a custom connector, following Anthropic's current guidance, or from ChatGPT if your plan supports custom MCP connectors.
7. Ask: "How did last week compare with the week before?" Check the tool call in n8n's executions.

#15Try it yourself

Expose one read-only tool through an MCP Server Trigger with Bearer auth. Connect it to one assistant and ask three questions. Check executions to confirm each call, then decide what a second tool should be.

#16Quick quiz

1. How do you expose your own workflow as an MCP tool?

Answer: Attach a Custom n8n Workflow Tool to the MCP Server Trigger.

2. Which URL works without publishing?

Answer: The test URL.

3. What should tools return?

Answer: Compact JSON with only what the assistant needs.

#17What should you check before the next lesson?

• The trigger has authentication set.
• Tools are read-only and return compact JSON.
• Calls are visible in executions.

Next: Error handling, retries and rate limits. Or return to the n8n for Marketers track.

#18Common questions

Do I need to self-host for MCP? No. The MCP Server Trigger works on Cloud with a public URL.

Can ChatGPT use my n8n tools? If your ChatGPT plan supports custom MCP connectors. Check OpenAI's current docs.

Is MCP secure? It's as secure as your auth and your tools. Use Bearer or OAuth and read-only tools first.