convez.ai
MCP Server
thousandflowers

skillreaper

thousandflowers/skillreaper

Ver no GitHub
BomPluginScore · de 0 a 100

Evidence-based pruning for your AI-agent stack — scan, report, and safely prune unused skills/MCP servers/agents using real transcript evidence

Instalação

$npx convez-ai thousandflowers/skillreaper

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
66
Forks
8
Acompanhando
1
Saúde
85%
Último commit
há 1 dia
Versão
v0.9.0
Issues abertas
12
Contribuidores
1

README

A tradução para o português deste README ainda está na fila.

brew install thousandflowers/tap/skillreaper
reap

One command. Zero config. Read-only. It scans your transcripts, finds every skill/agent/MCP your AI loads but never uses, and tells you exactly what it costs in context window, latency, and money.

The problem

Every Claude Code session loads 150–300 skill descriptions, agent configs, and rule files into context. Most of it is dead weight:

  • 187 items scanned
  • 142 never used (76 %)
  • 8 000 tok/session wasted
  • ~2 160 000 tok/month burned on irrelevant instructions

Your agent scrolls through a wall of irrelevant tools looking for the right one. Wrong picks cost turns. Turns cost tokens. Tokens cost money.

reap points at the waste. You decide what goes.

Before → After

Before skillreaperAfter skillreaper
187 items loaded every session45 items, all actively used
Wrong tool 1 in 5 turnsRight tool on first try
8 000 tok/session deadFull context budget for real work
~30 pages of irrelevant instructions read monthlyZero
Lower cache hit rate = higher latencySmaller prompt fits in cache

Install

# macOS — Homebrew
brew install thousandflowers/tap/skillreaper

# Any platform — Go (Go ≥ 1.22)
go install github.com/thousandflowers/skillreaper/cmd/reap@latest

Binary downloads — macOS (Intel + Apple Silicon), Linux (amd64 + arm64), Windows (amd64 + arm64) — all on the releases page. Single static binary, no dependencies.

Upgrading, uninstalling, and platform-specific tips → INSTALL.md.

Usage

reap                     # scan + report (read-only)
reap prune               # quarantine unused items (reversible)
reap keep <name>         # protect an item from pruning
reap restore --all       # undo every prune
reap --json              # structured JSON output
reap --md                # markdown report
reap --days 7            # shorter evidence window
reap version             # print version

Everything is reversible. reap prune moves files to a reaped/ directory with a versioned manifest. Nothing is ever deleted. Run reap restore --all and everything goes back exactly where it was.

Verdicts

LabelMeaning
REAPZero uses — safe to quarantine
KEEPUsed, tiny, or manually protected
REVIEWToo new or not enough sessions

Every verdict includes a reason suffix explaining why.

Privacy

100 % local. Zero telemetry, zero network, zero uploads. Reads config files and session transcripts on disk — your data never leaves your machine.

Platform support

PlatformFull support
Claude Code✅
OpenCode✅
Codex CLI✅
Hermes✅
CursorInventory only (no local transcripts)
OpenClawInventory only (no session history)

How it works

  1. Auto-detect — probes every known config directory. Only installed platforms are scanned. No flags needed.
  2. Inventory — scans skills, agents, MCP servers, hooks, and prose files across all detected platforms.
  3. Evidence — parses session transcripts (JSONL or SQLite). Counts tool_use blocks and command invocations with timestamps.
  4. Cost — character weight (ceil(chars / 3.7)) + init parser tool declarations. Model pricing auto-resolves by model name.
  5. Verdict — REAP / KEEP / REVIEW with machine-readable reason.
  6. Act — reap prune quarantines. reap restore --all undoes.

Limitations (transparency)

Token counts are approximate. The tool estimates tokens as ceil(chars / 3.7), based on the average English BPE tokenizer rate. Real token counts vary by tokenizer (Claude vs GPT vs Gemini) and content (more code ≈ more tokens per char). This is a documented approximation — the relative ranking matters more than the absolute number.

Platform format stability. Each supported platform has its own config layout and transcript format. These change over time as platforms evolve. Parser updates are an ongoing maintenance reality. The project is architected for easy fixes (one struct per platform in internal/platform/), but format changes can lag by days to weeks after a platform update.

Not a tool declaration fix. Claude Code's deferred tools reduce the init-time tool declaration overhead. Skillreaper addresses a different problem: always-loaded skill/agent/prose files. If a skill description is 248 characters, it is read into context every session — regardless of lazy tool loading. These two optimizations are complementary, not competing.

Design

  • 100 % local, zero dependencies, single static binary (Go ≥ 1.22)
  • Multi-platform — adding a new platform is one struct in internal/platform/
  • Reversible quarantine — never deletes, never destructive
  • MIT licensed
cmd/reap/       CLI entry point
internal/
  platform/     platform definitions + auto-detection
  scan/         inventory scanners
  usage/        transcript parser (JSONL + SQLite)
  report/       verdict logic + ANSI/JSON/MD renderers
  prune/        reversible quarantine
  cost/         model pricing
docs/           demo assets

Ver o original no GitHub

Bom para

Relacionados