Agent Token Audit Bench

See where each agent's tokens actually go, call by call

Problem

After switching to a heavily plugin-based agent, one user found long tasks cost more than double their previous setup — a $350 day with no visibility into which call, sub-agent, or tool round burned the money. Official dashboards show only a total.

Solution

A local proxy layer that logs each request's prompt, tool calls, sub-agent forks, and token counts, then aggregates them into a task-and-tool flame view of time and spend. It flags repeated calls and bloated context and suggests cuts.

Users

Heavy AI coding users running long tasks, indie developers and small teams juggling multiple agents

MVP

Local proxy intercepting and logging requests and tool calls; per-task/per-tool token and cost rollups; alerts for repeated calls and oversized context; exportable comparison reports

Revenue

Free or one-time local license; team subscription for aggregated comparison and alerting

Why now

Agentic coding has made token spend opaque while official billing stays aggregate-only, and a local sidecar has no platform dependency