SkillKit

Production observability for AI agent skills. Monitors usage, detects conflicts, analyzes cost, and prunes what you don't use.

Commands

SkillKit is a standalone CLI. Run with bunx @crafter/skillkit <command> or install globally (bun i -g @crafter/skillkit).

Usage & Budget

• skillkit scan - Discover skills + index session data (auto-runs on first use)
• skillkit scan --include-commands - Also track slash commands (not just skills)
• skillkit stats - Usage analytics with sparklines, streak count, weekly velocity (last 30 days)
• skillkit stats --all --days 90 - Full list over 90 days
• skillkit stats --json - JSON output
• skillkit list - Installed skills with size and context budget
• skillkit health - Unused skills + metadata budget check
• skillkit health --json - JSON output

Sessions & Activity

• skillkit sessions - Daily usage across all agents
• skillkit graph - 52-week contribution heatmap
• skillkit graph --json - JSON output
• skillkit auto - Auto-scan after Claude Code sessions (SessionEnd hook)

Cost Analysis

• skillkit context - Context tax: tokens + cost of CLAUDE.md, skills, memory per API call
• skillkit context --sonnet --turns 60 - Custom model pricing and session length (default: sonnet)
• skillkit context --opus / --haiku - Other model pricing
• skillkit context --json - JSON output
• skillkit burn - Subscription burn rate analysis (cost, models, daily)
• skillkit burn --days 7 - Custom range
• skillkit burn --plan 200 - Monthly plan cost in USD (default: 200)
• skillkit burn --json - JSON output

Quality & Conflicts

• skillkit conflicts - Detect trigger collisions between skills
• skillkit coverage <skill-path> - Find dead weight (unreferenced sections + files)

Cleanup

• skillkit prune - List unused skills
• skillkit prune --yes - Confirm deletion
• skillkit prune --skill <name> - Prune a single skill
• skillkit prune --yes --json - Confirm and output JSON

Trace (Internal)

• skillkit trace <prompt> - Record skill execution trace (powers conflicts + coverage)
• skillkit trace --list - List recent traces
• skillkit trace --list --json - List recent traces as JSON
• skillkit trace --show <id> - Show trace details
• skillkit trace --model <model> - Model to use for trace

Agent Filters

Any command accepts agent filters: --claude, --cursor, --codex, --gemini, --windsurf, --amp, --goose, --kiro, --roo, --opencode

When to Use

• "which skills do I use the most?" -> skillkit stats
• "are there unused skills?" -> skillkit health then skillkit prune
• "what's my context tax?" -> skillkit context
• "how much am I spending?" -> skillkit burn
• "do any skills conflict?" -> skillkit conflicts
• "is this skill bloated?" -> skillkit coverage <path>
• "show my activity heatmap" -> skillkit graph
• "daily usage breakdown" -> skillkit sessions

Decision Guide

1. First time? skillkit stats -- auto-discovers and indexes everything
2. Full picture? skillkit stats --all --days 90
3. Activity? skillkit sessions then skillkit graph
4. Context tax? skillkit context
5. Cost check? skillkit burn
6. Conflicts? skillkit conflicts
7. Dead weight? skillkit coverage ~/.claude/skills/my-skill/
8. Cleanup? skillkit health then skillkit prune --yes

How It Works

Discovers skills across 12 agents: Claude Code, Cursor, Codex, Gemini CLI, Windsurf, Amp, Continue, Goose, Kiro, Roo Code, Antigravity, OpenCode. Session connectors parse JSONL/JSON traces for Claude, Cursor, Codex, Gemini, and OpenCode. Deduplicates skills by name and inode. Self-healing DB deduplication on open. All data local in ~/.skillkit/analytics.db. Plan config stored in ~/.skillkit/config.json.