MCP Node Best Practices
What are MCP nodes?
With MCP nodes, you can prompt your way to a node that does exactly what you want it to do. They’re perfect when you want to interact with your favorite tools in ways that aren’t supported natively in Gumloop (in other words, when a node doesn’t exist for what you want to do) or when you want to run a more complex query, like chaining actions together or filtering.
Let's explore some tips and best practices for creating great MCP nodes.
Narrow prompts are best
To understand what makes a great prompt, think of a Slack MCP node like a sandwich shop — let’s call it the Slack Deli.
When you walk in, the deli offers a set menu: different types of bread, meats, cheeses, and toppings. You can mix and match what’s available, but you can’t order ingredients that aren’t in the shop. You can ask for no pickles or extra mustard, but you can’t get sushi or tiramisu — it’s just not part of what the Slack Deli serves.
That’s exactly how MCP nodes work.
When you ask Gummie to generate a Slack MCP node, it’s pulling from Slack’s set of available actions, not Gmail’s, not Salesforce’s, or any other tools. So your prompt needs to stay within Slack’s actual capabilities.
A great prompt is one that clearly describes what you want, using only what’s available on the menu.
✅ Here's a great prompt:
“Get all unread messages from the #sales channel in the past 3 days, sent by Jason.”
This works because it asks for message data, something Slack's MCP is able to handle.
❌ And a not so great prompt:
“Summarize conversations that involve an upcoming launch.”
This won’t work, because Slack (and its MCP tools) can retrieve messages, but not understand what “summarize” or “upcoming launch” mean — those aren’t ingredients on the Slack menu.
When prompting, you'll see example prompts and available MCP tools for that node to guide you.
Gummie is better the second time around
When testing your node with actual data for the first time, you may notice that it fails, which can be discouraging. All that work and the node doesn’t work?
The good news is that Gummie incorporates the errors in rewriting the code for the node. Its success rate on second tries is dramatically higher, as it better understands the actual inputs, parameters, and outputs you intend to use.
