MCP Audit

Analyze MCP server overhead and recommend cleanup.

Trigger

Use when:

• Sessions feel slow or expensive
• Adding a new MCP server
• Context fills up quickly
• Reviewing project configuration

Key Insight

Each MCP server adds ALL its tool descriptions to every API request. A server with 20 tools adds ~2K-4K tokens per request, regardless of whether you use those tools.

Audit Steps

Step 1: List Active Servers

Check all MCP configurations:

cat .claude/settings.json 2>/dev/null | grep -A 50 "mcpServers"
cat ~/.claude/settings.json 2>/dev/null | grep -A 50 "mcpServers"

Step 2: Count Tools Per Server

For each server, estimate token overhead:

• 1-5 tools: ~200-500 tokens (low overhead)
• 6-15 tools: ~500-1500 tokens (moderate)
• 16-30 tools: ~1500-3000 tokens (high)
• 30+ tools: ~3000+ tokens (excessive — consider tool filtering)

Step 3: Check Usage

Questions to ask:

• Which servers were actually used this session?
• Which servers haven't been used in 7+ days?
• Are there servers with overlapping functionality?
• Are there servers only needed for specific tasks?

Step 4: Recommend Actions

Disable servers that:

• Haven't been used in 7+ days
• Overlap with another active server
• Are project-specific but you're in a different project

Keep servers that:

• Are used every session (filesystem, git)
• Provide unique capabilities needed for current work
• Have low tool count (<5 tools)

Output

MCP AUDIT
  Active servers: [N]
  Total tools: [N]
  Estimated overhead: ~[N]K tokens per request

  Server Analysis:
    [name] — [N] tools, ~[N] tokens
      Status: KEEP / DISABLE / REVIEW
      Reason: [why]

  Recommendations:
    Disable: [list]
    Keep: [list]
    Review: [list]

  Projected savings: ~[N]K tokens per request (~$X.XX per session)

Thresholds

• Total servers: <10 (ideal), 10-15 (monitor), >15 (reduce)
• Total tools: <80 (ideal), 80-120 (monitor), >120 (reduce)
• Per-server: <15 tools (ok), 15-30 (filter), >30 (split or disable)

Rules

• Never disable servers without user confirmation
• Estimate token savings for each recommendation
• Consider task context — a server might be unused today but critical tomorrow
• Check for disabledMcpjsonServers to avoid re-recommending already-disabled servers