Dashboard
Overview
The first analytics tab: six KPIs, spend and block trends, anomalies, and which workflows are looping.
Overview is the default tab inside Dashboard. Its own subtitle asks the three questions it exists to answer: "Is your LLM spend safe? Are your agents looping? Is TokenPolice paying for itself?"
What it's for
A single screen you can scan in ten seconds each morning. Everything here respects the time-range control at the top, and every KPI compares against the previous period of equal length.
The KPI row
| Tile | What it counts | What it suggests |
|---|---|---|
| Sessions | Conversations in the range. | Traffic volume. A drop with steady spend means each conversation got more expensive. |
| Model Calls | Provider calls that ran. | Rising calls per session usually means agents are taking more turns. |
| Spend on Models | USD at provider list prices. | The headline number. |
| Blocked Calls | Calls actually stopped, enforced only. | Zero here with rules armed means nothing has hit a limit — or nothing is enforcing. |
| Savings if Enforced | Estimated spend a dry-run rule would have avoided. | Your case for promoting a rule to Enforce. |
| Model Reroutes | Swaps actually made, with the would-have count beside it. | Whether your reroute rules are firing at all. |
The savings tile is explicitly an approximation: "Dry-run rules check every call but let it through, so we can put a value on the spend enforcing them would have avoided. Treat the total as an approximation, not an exact figure."
The panels
| Panel | What it shows |
|---|---|
| Spend anomalies | A red strip, only when something spiked — the top three user / model / workflow combinations whose spend is far above their own recent baseline, with the actual and baseline amounts. |
| Allowed vs Blocked Calls | A stacked series: allowed calls, budget blocks, budget blocks if enforced, loop blocks, loop blocks if enforced. Enforced is solid, dry-run is hatched. A block rate sits on the right. |
| Spend Trend | This period solid, the previous period dashed, with the total and request count. |
| Unit Economics | Cache hit rate, dollars per session, runs per session, and blended dollars per 1K tokens — each with a delta. All at provider list prices. |
| Loop detection | Firings split by detector — exact prompt repeat and per-trace cap — with enforced and shadow counts and how many distinct runs were involved. |
| Loopiest workflows | Workflows ranked by loop impact, with the spend on looping runs. |
What to look for
- Spend rising while sessions are flat — cost per conversation is growing. Open Workflows and check prompt-size drift.
- A big "would-block" band and no blocks — your rules are watching but not acting. Check which switch is open on the Audit log.
- Loop firings on a workflow you thought was fine — open the run from Loopiest workflows and read its loop fingerprint.
- An anomaly naming one user — jump to Users and cap that pool.
- Everything reads a dash — nothing has been reported in this range. Widen the range, or finish setup.