MCP · Deterministic · VRE-backed semantic engine

Your agent re-spends the same tokens.
Now you can see the semantically-near waste too.

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.

Star on GitHub
278M
tokens in one day
our stack, measured
289K
tokens per request
a context you can't see
0.78%
cache-miss, still fresh
2.17M tkn/day ≈ $28/mo

Not another hash-based profiler

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.

⚡ Real semantic engine
Gematria token projection + fixed-parameter attention over every payload — not just a hash of the bytes. It sees near repeats, not only exact ones.
🧠 Trained latent adapter
A learned latent projector (embedded in the Pro backend) refines similarity in an adapted space — closer to how a human sees "the same command, reworded."
🔒 Deterministic, no hallucination
The delta is computed by a fixed engine — same input, same number, every time. No LLM guessing, no fabricated savings, no "it might be around…"
🕵️ Finds the invisible waste
The cache-miss tail and cross-model redundancy that provider caches can't share — the part of your bill that's real but invisible in every dashboard.
📊 Role-aware spend
System / user / assistant / tool, split out. See exactly where cost lives, not a single blob number.
🚀 Request-path proxy
Dedupe before billing. Snap a base_url at an OpenAI-compatible proxy and strip ~95% of re-sent input on a tool-heavy session.

Why this waste is invisible to you right now

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
tools
30% reuse
user
13% dup
assistant
11% dup
the harness finds
useful vs redundant

What the engine actually does

📊 Token spend by role

System / user / assistant / tool — exactly where the cost lives, at a glance.

🔁 Exact duplicates

The same tool result re-sent turn after turn, flagged with its cache key.

🧩 Semantic near-duplicates

Repeated content that is almost the same — caught by the gematria/attention engine, not by an n-gram guess.

🧠 System-prompt bloat

If system dominates, get a cache / summarization flag so you know what to fix first.

The "what we don't fake" contract

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:

✅ Deterministic by construction
Same payload in → same savings number out. Forward-reproducible, no stochastic LLM in the scoring path. Verify it.
✅ Opaque but auditable
The semantic method runs server-side and returns only numbers. Sensitive content is scrubbed before it leaves your machine. You see the contract, not the pearl.
✅ What it doesn't claim
It does not generate answers, write code, or "reason" — it measures. The claims here are about the numbers it produces, and those are testable.
✅ Real, measured base
Our 278M/day stat is our own stack's real export, not a marketing number. The 0.78% cache-miss tail is real and reproducible.

Start free, or run the full Pro engine

Free Trial
$0 · 1 hour
  • Complete Pro backend — no card
  • De-identified semantic-delta number
  • See exactly why the engine is worth paying for
Pro
$250 · 30 days
  • Real semantic engine + trained latent adapter
  • Semantic near-duplicate detection, not just exact
  • Server-side, de-identified end-to-end
  • Pay in crypto — no KYC, renew when you want
mcp-token-saver · deterministic · MIT core, crypto-native Pro
Compatible with any OpenAI-compatible agent. No-LLM scoring path — the numbers aren't hallucinated, they're computed. · GitHub