Testing out n8n-mcp · AI Automation Society

Testing out n8n-mcp

I took the wonderful n8n MCP https://github.com/czlonkowski/n8n-mcp for a test drive.

Obviously, I didn't start with a "Hello world" but rather with a "build a world!" example. Fail fast, that's my moto ;-)

I asked for a "workflow builder" workflow:

Get a webhook request where a users describes their desired automation
Use the above MCP to generate a workflow
Create the workflow on my server
Send notifications to mail and telegram

Observations:

Surprisingly Sonnet gave better results than Opus (think mode activated).
Both models initially chose to use code nodes instead of the built-in MCP node
Sonnet used the core n8n node in order to create the desired workflow; Opus chose a code node.

Asking for improvements, Sonnet added a un-connected MCP node, and Opus wasn't able to do so.

I hoped this approach would give me better results than the approach

@Nate Herk

shared with us recently, but so far it's not really the case.

Nevertheless, this is an exciting step forward. Within weeks from now we will probably all use an automation to bootstrap our workflows. Our job will be to be better at crafting prompts and in fixing and adapting the automated results to the specific use case and client.

Has anyone else here tried it and got better results?

Do you have a proven workflow you could share?

----

N.B. the *initial* prompt I used was:

```

Build a workflow that receives a request via a webhook and uses n8n-mcp (source: https://github.com/czlonkowski/n8n-mcp) to generate a n8n workflow and does the following things:

* Create a workflow on my n8n instance

* Sends a Telegram message notifying me that a user created a workflow

* Send an email back to the user, congratulating them for creating the automation, with a link back to the workflow

The webhook request will contain at least a user prompt with a description of the desired automation and an email to send the response back to.

```

1 comment