mcp-audit

Read-only inventory of MCP-server tool usage from Claude Code session transcripts.

When to run

• Before removing an MCP server (confirm zero use)
• Periodic cleanup (suggest monthly via plugin-audit cron)
• After adding a server, ~2 weeks in (validate it's earning its catalog spot)

How

bash ~/.claude/skills/mcp-audit/audit.sh [days]

days defaults to 30. Scans ~/.claude/projects/*/*.jsonl for mcp__<server>__<tool> invocations within the window, prints:

1. By server: count of tool calls per server, sorted desc
2. By tool: top 20 individual tools by call count
3. Zero-use servers: connected MCP servers (from claude mcp list) with no calls in window
4. Single-use servers: warning band (used 1–3 times — possibly tested-and-forgotten)

Output

Plain markdown to stdout. No mutations. No settings touched. Pipe to a file if you want history:

bash ~/.claude/skills/mcp-audit/audit.sh > ~/.claude/plans/mcp-audit-$(date +%Y-%m-%d).md

Acting on results

• Zero-use server, ≥30 days connected → strong remove candidate. Confirm with claude mcp remove <name> -s user.
• Single-use server, recent → leave in place (maybe newly added).
• High-use server with low-quality tools → consider whether to disable specific tools rather than the whole server.

Limits

• Counts tool invocations, not tokens or wall time. A single gh issue create call may be more valuable than 100 playwright_snapshot calls.
• Doesn't see deferred tools that were never ToolSearch-loaded. If a server's tools are lazy and never reached, it shows as zero-use even if relevant.
• Doesn't infer intent — a server may be unused because the workflow it supports hasn't come up recently, not because it's redundant.