Skip to main content
Run your node against the real service before you put it in a workflow. Studio has a Nodes tab for exactly that, and nothing has to be published first.

​
In Studio

Open the Nodes tab and pick your node. You get three panes. The node, as everyone else will read it: name, category, description, and the whenToUse you wrote. Any credential it needs is named here too, so a node asking for something you have not set up says so before you run it. The settings, rendered from your config.yaml. This is the same form Canvas shows, built from the same file, so a field that reads badly here reads badly everywhere. The output, empty until you press Run. Fill in the settings, press Run, and the node calls the real service. A streaming node streams into the pane as it arrives. This is the fastest way to see whether your node works, and it is also the honest preview of your config.yaml. If the labels are unclear or the fields are in a strange order, fix it now rather than after someone else meets it.

​
Load sample

If your node has test.yaml, Studio shows a Load sample button. It fills the form from testData.config and the inputs from testData.inputs, so you are one click from a run rather than typing settings each time. That is most of why test.yaml earns its place.
test.yaml
Write the fixture as a real request, not a minimal one. It is the sample every future reader loads first.

​
From the command line

The same run, without opening anything:
It prints the call it is about to make, the sample data it is using, what each output received, and whether the expect assertions passed.
The platform does not need to be running. It is your node, your machine, the real service.

​
Keys stay yours

Both routes need a real key, and neither stores one. The command line reads your own .env, using the credential name and field in upper snake case with any trailing Credential dropped. So openAICredential.apiKey is OPENAI_API_KEY. A missing one is named before anything runs.

​
What expect is for

expect turns a run into a check. Each key is an output, and each value is an expression over output that has to come back true.
Assert the shape, not the words. A model writes something different every time, so output.text.length > 0 holds and output.text === "Hello" does not. For a service node, name the method to call:

​
What a run cannot tell you

A node can pass here and still misbehave in a workflow, because the bench feeds it testData.inputs while a workflow feeds it whatever the edges carry. When that happens, the difference is almost always a template path: the node id or the output name does not match the edge you drew. Troubleshooting covers it.
Next: Discoverability