Claude Skill Audit: see what your Claude Code setup costs and what you can cut

Install a few plugins and connect a few services, and every session quietly starts heavier: less room for your conversation, and more context re-sent with every request. Claude Skill Audit answers three questions with evidence: what does a session load before my first message, which skills and MCP servers do I actually use, and how many tokens do I use, and where?

On one real setup it found a 39,000-token startup context, 96 skills installed but only 6 used in 45 sessions (about 8,100 tokens of skill listings every session), and MCP connectors that were connected but never called.

What’s inside: 3 skills, 1 SessionEnd hook, Python scripts. Published by: Rahul A P, the author of Zykit (open source, MIT), itsrahulap/claude-skill-audit.

How it works

Three skills, each backed by a Python script that reads what Claude Code already stores locally:

  • /context-audit reads the first request of recent sessions for the API-reported startup total, then estimates the split: tool schemas, system prompt, CLAUDE.md and memory, skill and agent listings, MCP instructions and hook-injected text. It also lists every MCP server, where it’s configured, whether it’s connected or needs auth, and whether you ever call it.
  • /skill-audit inventories your skills, commands and agents (your own, the project’s, plugins’ and claude.ai-synced ones), counts how often each was used, and estimates the listing tokens each source adds. --html writes a dashboard.
  • /token-audit sums API-reported usage by project, model, day and session, with the cache hit rate and a dated API-equivalent cost estimate.

Every number is labelled measured (reported by the API in your transcripts) or estimated (characters ÷ 4, or a price calculation). Claude Code deletes transcripts after 30 days, so a small SessionEnd hook appends a one-line summary of each finished session (about 450 bytes) to ~/.claude/skill-audit/history.jsonl, which keeps months of history.

Install

github.com/itsrahulap/claude-skill-audit

/plugin marketplace add itsrahulap/claude-skill-audit
/plugin install skill-audit@claude-skill-audit
claude plugin marketplace add itsrahulap/claude-skill-audit
claude plugin install skill-audit@claude-skill-audit

Skip the first line if you already added the claude-skill-audit marketplace. Reload the VS Code window or restart claude afterwards. New to plugins? See the install guide in this series.

How to use it

Type one of the commands, add options after it, or just ask in plain words. Each report ends with a short "Worth reviewing" list: what to look at, the evidence, the impact, what you’d lose and how to change it. The plugin never removes anything for you.

  • /context-audit
  • /context-audit --save-baseline
  • /context-audit --drift
  • /skill-audit
  • /skill-audit --unused
  • /skill-audit --html
  • /token-audit
  • Why is my context so big?
  • Which plugins can I remove?
  • How many tokens did I use this week?

Good use cases

  • Finding out why sessions feel heavy or reach compaction sooner than expected
  • Deciding which plugins, skills and MCP connectors earn their place, after installing some or once a quarter
  • Seeing which project or session dominates your usage, and whether prompt caching is working
  • Saving a baseline and checking what changed after you add or remove a plugin

When not to use it

If you run Claude Code with only a couple of plugins and no MCP connectors, there is little to find. On a subscription, the cost figures are curiosity rather than money saved; the context numbers still matter.