Measured against caveman & ponytail — and wins
Cut Claude Code's token bill on three axes
A terse dev persona, a tool-output compressor, and context-diet rules — one plugin. Across 20 live tasks it billed 52% of a bare model, with 1 backfire instead of 6 or 8.
or claude plugin marketplace add JayPokale/RDXmin
52%
of a bare model's 20-task bill
1/20
backfires — rivals hit 6 and 8
~46%
shrink on oversized tool output
0
dependencies, network calls, LLM calls
One plugin, three levers
Other token savers only make the model write less. Tool output is the bigger bill — 67.5% of a session — and it re-bills every turn.
How the model writes
Terse persona
Senior-dev voice: fragments over sentences, YAGNI-first code, reuse before new code. Three levels — /rdx lite, full, ultra. Commits and security warnings stay verbose on purpose.
What survives into context
Output compressor
A PostToolUse hook shrinks tool results before the model reads them: ANSI scrub, head + tail elide with error-line salvage, same-session dedup. Never touches Read, Edit, or Write.
What gets read at all
Context diet
Rules that teach the model to fetch the slice, not the file: grep first, sliced reads, filter at the source, never re-read what's already in context.
Live savings statusline
A ⇣9k tok badge showing measured chars elided — real baseline, not an estimate.
Works beyond Claude Code
Generated rulesets for Cursor, Windsurf, Cline, Kiro, and Copilot ship in the same install.
Tested where it matters
The compressor is where a bug corrupts files — it's covered by the test suite, with a hard allowlist.
Easy off-switch
"stop rdx" for the persona, RDX_COMPRESS=0 for the hook, npx rdxmin --uninstall for everything.
400 lines of output. Two matter.
Tool results re-bill on every later turn. The compressor keeps the head, the tail, and the error lines — the rest never reaches the model.
Deterministic — no LLM calls, no network, no dependencies. Allowlist: Bash, Agent, WebFetch, WebSearch, Grep, Glob, mcp__* — never Read/Edit/Write.
The 20-task bill
Four arms, same 20 live tasks, same model, isolated configs, billed tokens — shown as a share of the bare model. Raw transcripts and the runner ship in the repo.
Average task: 69% vs 91% / 98%. Worst single task: 173% vs 227% / 424%. Both rivals are credited in the repo's prior-art table — right above these numbers.
Built in the open
MIT licensed, zero dependencies. Land a PR — or file an issue good enough to fix itself — and you appear here automatically.
Co-engineered with Claude, Codex & Antigravity · descended from caveman & ponytail
FAQ
No — terse ≠ incomplete. The rules cut words around the facts, never the facts: the fix, the gotcha, the caveat all stay. In the 20-task benchmark every RDXmin answer was verified correct.