Observability

Every run - from the playground, the API, a schedule, a channel, or an eval - is persisted with its inputs, its output, and the ordered sequence of tool calls and results that produced it.

Traces

A trace is the run’s steps in order: each tool call with its arguments, and the result that came back. It is the same data live and after the fact, so a run you watched stream is replayable later without extra instrumentation.

cruqai runs                 # recent runs
cruqai trace <run-id>       # one run's persisted trace

In the dashboard this is AI Monitoring.

Metering

Each run records prompt tokens, completion tokens, total tokens, cost in USD, and wall-clock duration, taken from the provider’s own usage block.

run       agent     status     created           cost  time  input
1ee3a5fe  69f7142c  completed  2026-08-06 12:05  $0.0031  2.4s  summarise Q3

Cost is exact decimal and is carried as a string on the wire - JSON floats cannot represent it faithfully. Every metering field is nullable, and null means unknown. Some providers report no usage block; that is not the same as a run that cost nothing. The CLI prints - rather than $0 so the two cannot be confused, and anything summing these values can tell them apart.

Eval runs carry source = "eval" and are excluded from AI Monitoring and cruqai runs by default. A single model comparison can add hundreds of rows, which would otherwise bury real traffic.

Over the API

GET /v1/runs?agent=<id>&limit=20
GET /v1/runs/{id}          # the run plus its steps

Usage is nested under usage on each run:

{
  "id": "...",
  "status": "completed",
  "model": "openrouter/openai/gpt-4o-mini",
  "usage": {
    "prompt_tokens": 100,
    "completion_tokens": 50,
    "total_tokens": 150,
    "cost_usd": "0.000042",
    "duration_ms": 820
  }
}

Getting Started

Concepts

Reference

Traces

Metering

Over the API

​Traces

​Metering

​Over the API

Traces

Metering

Over the API