TokenMasterX — AI skill for Claude Code

TokenMaster

**A new cost model for code-understanding agents.** TokenMaster re-engineers *token economics at the harness layer* — the layer that decides what the model re-reads on every single turn — for [Claude Code](https://docs.claude.com/en/docs/claude-code) and GitHub Copilot CLI.

The thesis in one line: **the model should pay once to understand a codebase's structure, then never again.** Today's harnesses violate this on every turn — they hand the model a growing transcript and let it grep its way to understanding, re-paying for the entire accumulated context turn after turn. TokenMaster makes that re-derivation *economically illegal*: structural questions get routed to a **prebuilt code graph** answered in one bounded query, so the cumulative token bill collapses instead of compounding.

It attacks the bill on **two complementary fronts**: a **routing layer** (below) that collapses the cost of re-deriving structure, and **[Brainspace](#brainspace--the-compression-layer)**, a compression layer that shrinks the raw tool outputs sitting in the transcript. Different token populations, so the wins compound. Routing is the headline; Brainspace is the multiplier.

Baseline vs TokenMaster, measured

Same CLI. Same model. Same task. The only variable is whether TokenMaster's routing agent is on.

**-73% input tokens. 3.71x more efficient overall. Up to 7.8x on blast-radius analysis.** **12 / 12 tasks answered from the graph. Zero correctness regressions.**

*Pooled across scikit-learn + sympy. 36 live GitHub Copilot runs.* [*Full breakdown below.*](#by-the-numbers)

/token-master

That one command builds the index for the current repository and turns on routing.

---

Why token economics, not token *count*

Start with the one fact everyt