Every tool call grows the context. Every turn re-sends it. You pay for all of it — and most profilers only catch byte-identical repeats.
mcp-token-saver runs a deterministic semantic engine (gematria projection + fixed-parameter attention + a trained latent adapter) to find redundant tokens that exact-dedupe can't — the near-repeats. We measured our own stack: 962 requests, 278M tokens, $5.61 — in one day, on one model.
Every other token profiler stops at byte-identical duplicates. That misses the waste that actually compounds in agent loops: tool results re-sent with slight edits, log tails that differ by a line, answers re-spelled with a new timestamp. Those are semantically-near duplicate — and exact-dedupe can't see them.
A real 50-turn agent loop re-injects the same big tool results over and over. The striped area is what mcp-token-saver flags — including the part exact-dedupe misses.
System / user / assistant / tool — exactly where the cost lives, at a glance.
The same tool result re-sent turn after turn, flagged with its cache key.
Repeated content that is almost the same — caught by the gematria/attention engine, not by an n-gram guess.
If system dominates, get a cache / summarization flag so you know what to fix first.
This is not a black-box SaaS. The MIT core is readable, and the Pro semantic delta is computed the same every time. Here's the honest boundary — you can verify all of it: