Skip to main content
Tools and models are first-class analytical dimensions. Every tool call and every model step is captured with its arguments, results, latency, and cost, so you can rank, filter, and investigate at the tool or model level.

​
Tools

The Context Company records every tool invocation the agent makes: the tool name, the arguments passed, the result returned, whether it succeeded or errored, and its latency.

​
What you can analyze

  • Which tools are called most frequently.
  • Which tools have the highest error rate.
  • Which recurring failures involve a particular tool.
  • Which arguments were passed to a specific tool over a time range.
  • What results the tool returned, including for failing runs.
  • Which tools contribute most to latency.
  • Which tool sequences commonly precede an error or negative feedback.

​
API and MCP

Returns a breakdown of tool usage. See the API reference. MCP calls like get_agent_runs accept failureContains and onlyErrored, so you can pull runs where a specific tool failed directly into your coding agent’s context.

​
Example questions

Which tool is responsible for the most failed runs in the last week?
Find runs where the search tool returned an empty result and the agent still produced a confident answer.
What arguments were passed to issueRefund in the runs that errored?

​
Models

Model steps record which model was used, the input and output tokens, the resulting cost, latency, and any error at the step level.

​
What you can analyze

  • Models used, aggregated by run count, tokens, and cost.
  • Cost and token usage per model over time.
  • Latency per model.
  • Error rate per model where measurable.
  • Comparisons between models on the same task.
  • Filter any other analysis to a specific model.

​
API

Returns a breakdown of model usage. See the API reference.

​
Example questions

How have my costs shifted between models this month?
Which models are slowest on multi-turn conversations over 5 messages?
Did failure rate increase after we switched to gpt-4o-mini on the summarize step?
Because tool and model dimensions are indexed as native columns, Insight Search can join them freely with any other dimension:
Find runs where the search tool was called after the summarize tool, on gpt-4o, for organizations on the enterprise plan, where the final response didn’t answer the user’s question.

​
Limitations

  • Arguments and results are captured as recorded by the instrumentation. If your framework does not expose a tool’s arguments (rare), that field will be empty for those calls.
  • Model-level cost depends on the framework surfacing token counts. All supported frameworks do; custom instrumentation is responsible for reporting tokens.

Traces

See tool and model behavior inside a single run.

Failures

Runtime errors grouped, including tool-call errors.

Insight Search

Ask cross-dimensional questions about tools and models.

What TCC captures

The exact fields recorded per tool call and model step.