What is skillreaper?

Skillreaper is a command-line tool written in Go that helps developers clean up bloated AI coding assistant configurations. When you use tools like Claude Code, Cursor, or Codex CLI, those tools load a list of available skills, agents, and connected servers at the start of every session. Over time, developers accumulate dozens or hundreds of these items, most of which they never actually use. Skillreaper scans your local configuration files and session history, then shows you exactly which items are dead weight. The tool works entirely on your own machine. It reads configuration files and session transcript files that your AI coding tool stores locally, counts how often each skill or agent was actually invoked, and assigns each one a verdict: REAP means it was never used and is safe to remove, KEEP means it was used or is small enough to ignore, and REVIEW means there is not enough session history to be sure yet. Running "reap" by itself gives you the full report. Running "reap gap" shows a breakdown of how much of what gets loaded each session is ever actually fired, expressed as a utilization percentage and estimated token count. If you decide to act on the report, "reap prune" moves the unused items into a separate "reaped" folder rather than deleting them. A manifest file records exactly where everything came from, so "reap restore --all" puts everything back exactly where it was. Nothing is ever permanently deleted. The practical benefit is that a leaner configuration gives the AI fewer irrelevant instructions to read through at the start of each session. The README gives a concrete example: 187 items loaded per session but only 4 actually used, wasting roughly 8,000 tokens per session on instructions the AI never needed. Reducing that clutter can improve the accuracy of the AI's tool choices and lower the cost of each session. Skillreaper works with Claude Code, OpenCode, Codex CLI, and Hermes with full support, and can inventory (but not analyze transcripts for) Cursor and OpenClaw. Token counts are estimated rather than exact, and the tool is clear about that limitation. It is MIT-licensed and ships as a single static binary for macOS, Linux, and Windows.