SkillScope

You have 583 skills, subagents and hooks installed. 28 have ever fired.

SkillScope reads your local Claude Code session transcripts and tells you which skills, subagents and hooks actually run, what they cost in tokens, and which are dead weight.

It runs entirely on your machine. Nothing is uploaded. Nothing in ~/.claude is modified.

$ npx cc-skillscope

SkillScope
172 sessions · 27 projects · 114,965 transcript lines

Fired at least once: 28 of 583 installed
Never fired: 555  (skillscope skills --dead)

Top skills
  component                                   fires  sessions     tokens  last used
  caveman:caveman                                46        44  5,120,956  2026-07-21
  andrej-karpathy-skills:karpathy-guidelines     36        35  1,240,452  2026-07-21
  ponytail:ponytail                              28        27    924,484  2026-07-21
  superpowers:executing-plans                     4         4  1,631,399  2026-07-21
  frontend-design:frontend-design                 4         4  1,087,095  2026-07-16

Tokens
  measured total     171,234,864
  unattributed       124,622,130 (no skill was active)

A recorded demo GIF goes here before the v0.1.0 announcement.

npx cc-skillscope          # no install
npm i -g cc-skillscope     # or keep it around

Requires Node 18 or newer. The CLI is named skillscope; the npm package is cc-skillscope because skillscope was taken.

skillscope                 # summary (default)
skillscope skills          # every skill that fired
skillscope skills --dead   # every skill that never fired
skillscope agents          # subagents
skillscope cost            # measured tokens per component
skillscope wrapped         # shareable SVG stats card (--month 2026-07, --theme light)
skillscope doctor          # sanity-check installed skills, agents and plugins

Every command accepts:

wrapped also takes --month <YYYY-MM>, --all-time (default), --theme dark|light and --out <file>. doctor is read-only: it prints findings and the exact fix, and never modifies anything under ~/.claude.

Set CLAUDE_CONFIG_DIR to analyze a Claude Code directory somewhere other than ~/.claude.

Generate a shareable SVG stats card:

skillscope wrapped --theme dark --out docs/img/card-dark.svg
skillscope wrapped --theme light --out docs/img/card-light.svg

• Nothing leaves your machine. There is no telemetry, no network code, and no API calls anywhere in this package.
• Transcripts are read, never written. The only thing SkillScope opens for writing is stdout.
• dependencies is empty. Argument parsing is node:util.parseArgs; colour is a dozen ANSI escapes. Nothing else runs when you npx this.

Fires are counted, not guessed. A skill counts as fired when it is invoked by the Skill tool or by a /slash-command. A subagent counts when the Agent tool launches it. A hook counts once per entry in a stop_hook_summary record.

Direct tokens are measured. Claude Code stamps assistant turns with the skill that was active (attributionSkill), so per-skill tokens are read out of the transcript rather than inferred. Subagents write their own transcripts under <session>/subagents/agent-<id>.jsonl; those tokens are charged to the Agent call that launched them, matched through the tool-use id. Turns with no skill active are reported as unattributed rather than spread across components — that number is large, and pretending otherwise would make every per-skill figure look bigger than it is.

Overhead is estimated, and labelled as such. Every installed skill and subagent puts its description into the system prompt of every session. No transcript records that, so it is sized as description length / 4 × sessions. Treat it as an order of magnitude, not a measurement.

"Dead weight" is about components, not about you. Installing something you have not needed yet is not a mistake; it is just information worth having.

Fired but not installed is a third category: builtin agents like Explore, or skills from a plugin you have since removed. They are listed separately and never counted as dead.

Transcript formats change. Parsing lives behind src/parser/adapters/, so a new format is a new adapter file rather than a rewrite — see CONTRIBUTING.md.

• v0.1 — report, skills, agents, cost. (this release)
• v0.2 — skillscope wrapped: a shareable stats card, with themes.
• v0.3 — skillscope doctor: malformed frontmatter, shadowed names, orphaned plugin components, plus caching for large histories.

Good first issues are labelled. Fixtures live in test/fixtures/ and are the parser's contract: sanitized real transcripts with every piece of prose replaced. See CONTRIBUTING.md.

MIT