Skip to main content
is_empty marks a call that ran, succeeded, and gave the model nothing it could use. These are the failures nobody reports. The protocol calls them success, the client sees no error, the model gets an empty box, and you never hear about it. In the outcome breakdown they get their own slice for exactly that reason.

​
What counts as empty

The SDK inspects the result and marks it empty when it is:
  • an empty array
  • an empty object
  • an empty string
  • an array whose single text item parses to an empty array
That last case is the one worth knowing about. MCP results are usually a content array holding a text item, and the text is very often JSON — so a tool that “returns no rows” arrives as a populated content array containing the string "[]". Without unwrapping it, every no-results call would look like a healthy answer. An error is not also an absence, so is_empty is only ever set alongside an ok outcome.

​
Where it appears

​
An empty result is not always a bug

Sometimes nothing genuinely matched. A search over a small dataset returns nothing fairly often, and that is honest. The threshold is set at 5% of calls because below that it is ordinary and above it something is usually wrong in one of three ways:
  1. The model is asking for the wrong thing. Your description implies the tool covers more than it does.
  2. The arguments do not do what they appear to. A filter that silently ANDs everything, a date field expecting a format the model does not produce.
  3. The tool should have errored. An id that does not exist is a tool_error, not an empty list — and telling the model so lets it recover instead of guessing.
That third one is the most common and the most valuable to fix. An empty array tells a model nothing; isError: true with “no customer with that id” tells it exactly what to do next.