Overview
Quarterly audit pattern for MCP-using agents (Claude Code, Hermes, OpenClaw). Drops bloat, replaces narrow web/scrape skills with one Scavio MCP, re-measures token weight.
Trigger
Quarterly (manual)
Schedule
Quarterly
Workflow Steps
Log every tool invocation for 2 weeks
From session traces if available; honest journal otherwise.
Tag tools as used / unused / duplicate
Bottom 20% by usage = drop candidates.
Drop never-invoked skills
If you didn't use it in 2 weeks, you don't need it loaded.
Drop duplicate skills
Multiple 'fetch URL' / 'search' variants → keep one.
Replace 5-8 narrow web/scrape skills with one Scavio MCP
claude mcp add scavio https://mcp.scavio.dev/mcp.
Re-measure per-message input token cost
Before vs after; expect 4-8K input tokens saved per message at heavy usage.
Python Implementation
# This is a config-discipline workflow. The actionable part:
# claude mcp list # list current MCPs
# claude mcp remove <unused>
# claude mcp add scavio https://mcp.scavio.dev/mcp --header 'x-api-key: $SCAVIO_API_KEY'JavaScript Implementation
// CLI discipline; no application code.Platforms Used
Web search with knowledge graph, PAA, and AI overviews