What problem does it solve? Every session, your agent silently loads CLAUDE.md files, MCP server tool schemas, SessionStart hooks, skill descriptions, agent definitions, and memory indexes before you type a word — and each one costs tokens on every session. This Skill inventories those always-on sources, estimates their cost, and produces a prioritized, quality-preserving plan to trim them. ## Core Features & Use Cases - Read-only inventory and ranking: Sizes CLAUDE.md chains, MCP servers, hooks, skills, agents, plugins, and memory using a bytes÷4 token estimate, then ranks sinks biggest-first with quality-risk annotations. - Top-3 quick wins with concrete actions: Proposes the highest-reclaim, lowest-risk cuts (disable unused eager MCP servers, gate SessionStart knowledge dumps, dedup CLAUDE.md) with exact commands and estimated tokens reclaimed. - Safe, reversible changes: Proposes rather than changes by default; on approval or --apply, it disables or backs up instead of deleting. - Use Case: Your Claude Code sessions feel bloated and slow. Run the audit to discover an unused eager MCP server consuming ~4,200 tokens per session and a SessionStart hook dumping platform docs you never use, then reclaim ~60% of your baseline. ## Quick Start Ask the agent to run /token-audit to inventory your setup and show the top token-saving opportunities.
