Skip to main content
TokenOps answers what you spent this month, which agents drove it, and whether it’s going up. Everything is scoped to a period — the current month plus the previous five.

​
Summary strip

The gap between Active agents and total configured is often the most useful number — deployed-but-idle agents show up here.

​
Agent cost

Three controls above it:
  • Search — filter by agent name
  • Period — current month or any of the previous five; changing it reloads the summary strip too
  • Sort — Most tokens, Highest cost, Most operations, Slowest, or Name
Export downloads tokenops.csv reflecting your current search and sort — narrow to one team’s agents and export just those.

​
Token usage

A second strip splits totals four ways: Total tokens, Prompt tokens, Completion tokens, and Average tokens / operation. The prompt/completion split is the practical cost lever. A high prompt share usually means context is being resent — a prompt-engineering problem. A high completion share means verbose outputs — a model or instruction problem.
The cost of the platform choosing an agent is metered as its own line item, so routing never inflates an agent’s numbers. See TokenOps.

​
Where the numbers come from

Token counts are collected automatically from every LLM call — no instrumentation is added to your agents. Costs apply those counts to a per-model price table, versioned by date so a provider’s price change doesn’t re-cost last month. See observability.
For an LLM-generated summary of what changed rather than raw numbers, the finops/insights API returns exactly that — useful for a weekly digest. See TokenOps.

TokenOps reference

How token counts become dollar amounts.

Observability

Drill from a cost number into the traces behind it.

LLM router

Change which models your agents use.

Agents

Stop or scale down expensive agents.