The inspector tests configured operations. The chat playground can test how a model chooses and combines tools. These checks answer different questions and neither should be treated as a harmless preview of a live write.
Prepare safe test data#
Use an agreed test record, dedicated test account or fictional input that cannot cause an unintended business action. Identify whether the selected tool reads, writes, sends or spends. An administrator test can execute the real operation.
Test a tool directly#
Open the intended tool's test panel and enter a small valid input. Check the target account, HTTP result, returned fields and final MCP output. Also try a missing required input, invalid value and empty result where appropriate.
Do not share raw diagnostic output without inspection. Detailed request headers or bodies may reveal secrets or private records even when the normal configuration screen masks them. Use an approved administrator route for diagnostics.
Test the chat workflow#
If the playground is configured, choose the approved model connection and a narrow tool selection. Provider access and model usage may have separate cost and data-processing implications. Use the agreed account and budget; opening a playground does not make model calls free.
Ask for the same read-only task a customer will use. Inspect which tool the model selected, whether its arguments were correct and whether it grounded the answer in returned evidence.
Repeat through the real AI client#
An administrator test can use different credentials from a production client. Finish with the assigned public HTTPS connection: authenticate, discover the intended tools and run the agreed operation. Test missing/wrong credentials and account boundaries through that path.
Replaying a call#
Treat replay as another execution. Do not replay a message, order, publish or deletion operation until its effect and authorisation are clear. After a timeout, inspect the destination state before deciding whether another execution is safe.