Successful MCP calls can still leave the answer unclear

An MCP usage call succeeded, but the token counts in its response didn’t add up. The prompt and completion fields accounted for only part of the total. I could reconcile the numbers using the model responses, which counted reasoning tokens as part of completion. The usage response used the same term for visible completion and gave the agent no explanation for the difference.

That left me with a concrete way to check the tool. Could an agent explain the total using the response we actually gave it?

Defining the tasks

Earlier logs contained successful usage reads and sequences of route deployment and model invocation. They did not preserve the operators' questions, their final answers or any human help between calls. The indexed fields also could not establish that a deployed route was the resource subsequently invoked.

The review used a few small, controlled tasks in a dedicated test account. One reconciled existing usage across summary, daily and model views. Another requested structured model output and checked its trace. The third created a route, simulated it, deployed it and confirmed a real invocation against the deployed version.

Each task had a defined expected result. I handled login, organisation selection and consent. The agent needed no subsequent human correction or manual platform operation. It used a temporary HTTP client against the real server, with the live catalogue and recipes available. This did not test the native OAuth behaviour or recovery of a released coding client.

The tasks completed. They establish results for those checks in that account; the sample cannot support a general completion rate.

Reconciling the token counts

The following numbers are an illustrative example, not figures from the review. Suppose the model responses report this breakdown:

Token categoryExample count
Prompt100
Completion, including reasoning400
Reasoning within completion360
Visible completion40
Total500

A usage tool counting only visible completion could return:

{
  "prompt_tokens": 100,
  "completion_tokens": 40,
  "total_tokens": 500
}

This recreates the terminology problem I encountered: completion_tokens meant visible completion in the usage response, while the model responses included reasoning. The tool supplied neither a reasoning subtotal nor an explanation of the difference.

In the illustrative example, 100 + 40 + 360 = 500. The original review also reconciled once reasoning was included; it did not indicate incorrect billing. The issue was the use of the same term for different quantities. An agent relying on the usage response alone would need another source to explain the difference.

The review established that ambiguity. It did not establish that an agent had given a user a wrong answer because of it. A task-level check would need to inspect the explanation the agent produced.

Checking the relevant result

Cleanup provided another example of why status codes require context. A read after cleanup returned HTTP 200 with a 404 tool error inside. That 404 confirmed the expected result. The temporary route was gone.

The route task required a different check. Simulation exercised the draft path without making its final model invocation. After deployment, the review confirmed the active version and made a real call. The response identified the route and version, returned the expected output and supplied runtime confirmation.

Those observations concern different stages of the operation. Receiving a request, applying a change and confirming the intended result each need their own evidence. A later read helps only when it can be associated with the relevant resource.

Test the explanation a reader needs

I’d change the usage contract to say which tokens each field includes. If the underlying figures support it, I’d expose a separate reasoning subtotal. A follow-up review still found the distinction missing from the contract, so changes to telemetry elsewhere gave me no basis to expect a clearer answer.

The check I’d run after the change is the one the successful tool call left open. Ask an agent to reconcile the usage for the same project and inspect its answer. It should account for reasoning from the fields it can read or say which part of the breakdown the tool hasn’t supplied.