AI Agent Profiler — Measure agent cost, cache waste, and context bloat

I built a local-first profiler that sits as a transparent reverse proxy between your coding agent (Claude Code, OpenCode) and the LLM provider, recording every request without adding latency. It's like perf for your agent — showing you exactly where your tokens go.

I'd love feedback on

• What profiling blind spots am I missing? What can't you answer about your agent today?


• Has anyone tried proxy-level optimization and hit different economics? Did I miss a trick, or is the provider cache truly near-optimal for how agents send requests?


• Does knowing ~40% of your API bill is non-user overhead change how you configure subagents or tool briefs?


• Is anyone else bothered by the 5-minute cache TTL? What practical TTL do you observe in the wild?


• Shell hooks vs structured output — are AI-native tools with token-efficient structured output the right fix, or is output filtering good enough?