> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zelto.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Tool calls

> Track tool usage, failures, success rates, and duration across your voice agents.

The **Tool calls** page shows how your agents use functions across all their
calls. Open **Observability → Tool calls** to see which tools run most often,
which ones fail, and the calls behind each result.

## Overview

The summary reports:

* Total tool invocations in the selected period.
* Number of unique tools invoked.
* Success rate across invocations with a known success or error outcome.
* Failed invocations.

The activity chart plots daily tool-call volume and failures. Use the agent,
status, time-window, and tool-name filters to narrow both the summary and the
table. Filters stay in the URL, so a filtered view can be shared.

## Tool health table

Each row represents one observed tool and includes its invocation count,
number of calls affected, success rate, failures, P90 duration, and most recent
invocation. Select a row to inspect its latest invocations without leaving the
list.

The invocation panel shows the agent, provider, outcome, arguments, result,
error message, and a link to the original [conversation](/docs/conversations).
Arguments and results are collapsed by default so large payloads do not crowd
the timeline.

<Note>
  P90 duration is available only when the agent exports timed OpenTelemetry tool
  spans. Provider history without duration data shows no duration instead of a
  misleading zero.
</Note>

## Outcome handling

Zelto classifies a tool invocation as:

| Outcome | Meaning |
| - | - |
| Success | The provider reported a successful, completed, or OK outcome and no error message. |
| Error | The provider reported an error or failure, or included an error message. |
| Unknown | The provider did not supply enough status information to classify the invocation. |

Unknown outcomes are counted in total usage but excluded from the success-rate
denominator. This prevents missing provider metadata from being treated as a
successful execution.

## Related

* [Conversations](/docs/conversations#tool-calls): each invocation appears inline
  with the transcript.
* [Traces](/docs/traces): tool spans appear in the full OpenTelemetry waterfall.
* [Tool-check monitors](/docs/monitors#tool-checks): continuously verify that an
  expected tool executed correctly on each call.