Developer cuts Claude Code token costs by 77% using a transparent proxy optimizer

A developer built a tool called AI Agent Profiler (aap) to intercept and analyze every HTTP request made by Claude Code, an AI coding agent. The investigation revealed that agents re-send the entire conversation history on every turn, causing token counts to snowball — one 70-request session consumed roughly 3.9 million input tokens. The optimizer trims this overhead by replacing stale file reads with short stubs and dropping schemas for tools that were never called, accounting for the bulk of the savings. Anthropic's prompt caching already reduces costs significantly, but the tool ensures cache breakpoints are re-anchored after each trimming pass to maintain cache hits. The developer notes these are early results from a single test fixture and that individual percentages should be treated as approximate.