Cost & Usage
How the cockpit tracks tokens, cost, run limits, and surfaces warnings before you hit a wall.
The CLI tracks usage and cost as structured events flow in, then surfaces warnings when configured thresholds are crossed.
Where you see it
| Surface | What it shows |
|---|---|
| Header | The active model and (when available) abbreviated usage |
Cost drawer (/cost) | Per-agent and per-run breakdown: input/output tokens, cost, projection, model |
| Priority alerts strip | Warnings for crossed thresholds and approaching run limits |
What gets tracked
- Tokens — input and output, per agent and per run.
- Cost — derived from token counts and the model's pricing where known.
- Run limits — configurable caps that, when crossed, can trigger a pause or warning depending on policy.
Numbers come from structured usage_* events emitted by core. The CLI does not estimate or extrapolate beyond what it's told.
Thresholds and crossings
Thresholds are evaluated against incoming usage events. Crossing a threshold is itself an event that:
- updates the Cost drawer
- adds a row to the Priority alerts strip
- surfaces in the Timeline
Typical thresholds:
| Type | Example |
|---|---|
| Soft warning | Approaching 80% of a run's token budget |
| Hard warning | Crossed the configured cost ceiling |
| Run limit | Maximum number of agent steps reached |
What the cockpit does at a threshold
By default, threshold crossings are informational — you see them, you decide. If your permission mode and rules are configured to gate on run_command or change_model, the agent's next action may also surface in the Approval drawer.
The cockpit never silently kills a run for crossing a soft threshold.
Read next
- Drawers — the Cost drawer
- Concepts → Approvals — gating on actions
- Export — capture cost data with the run artifact