MCP Security Auditor

AI-consumed reference. Optimized for Claude to read during execution. Human-readable explanation: see docs/architecture/HIERARCHICAL_PLANNING.md or docs/getting-started/ depending on topic.

MCP Security Auditor

STATUS — v3.7.0-rc.1. Read-only auditor; does not enforce — that's mcp-call-gate.cjs's job.

Behavior

1. Read <security-dir>/mcp-audit.jsonl (append-only; produced by hooks/mcp-call-gate.cjs). Resolve <security-dir> the same way the writer does — see hooks/lib/security-dir.cjs: $AF_SECURITY_DIR, else .claude/security/ if it exists, else .aura/security/. Do NOT hardcode one of them: read the first that exists, and report "no audit log yet" if neither does.
2. Group entries by:
   • Agent → MCP server → method
   • Time bucket (last 1h / 24h / session)
   • Status (success / blocked / rate-limited)
3. Project to TOON (via scripts/json-to-toon.cjs --schema generic or custom fields) — NEVER load raw JSONL into context
4. Surface findings categorized as:
   • Blocked calls (BLOCKED:true) — deserve investigation
   • Rate-limit warnings — soft (80%) or hard (100%) hits
   • Suspicious patterns — destructive SQL detected, large output (>10KB), high frequency

What this skill does NOT do

• Does NOT enforce — mcp-call-gate.cjs is the enforcement layer
• Does NOT mutate the audit log (append-only by design)
• Does NOT call MCPs itself (read-only on the audit file)
• Does NOT proxy MCP calls — the gate hook intercepts at PreToolUse
• Does NOT decide policy — rules/agent/mcp-security-policy.md is authoritative

Audit log schema (per spec §23.2)

{"ts":"2026-05-07T...","agent":"architect","mcp":"context7","method":"query","input":{"q":"React hooks"},"output_size_bytes":4521,"latency_ms":234,"success":true}
{"ts":"2026-05-07T...","agent":"backend","mcp":"postgres","method":"query","input":{"sql":"DROP TABLE users"},"BLOCKED":true,"reason":"destructive_op_blocked"}
{"ts":"2026-05-07T...","agent":"frontend","mcp":"figma","method":"download_image","input":{"url":"..."},"success":true,"rate_limit":"80%_warn"}

Sanitized input fields per scripts/security/sanitize-mcp-input.sh:

• Authorization headers stripped
• Tokens/secrets replaced with [REDACTED]
• Content-Length > 1KB → truncated with ... [N more bytes]

Output (TOON)

mcp_audit_summary{window,total_calls,blocked,rate_limit_hits,top_agent,top_mcp}:
  24h,142,3,2,architect,context7

blocked[3]{ts,agent,mcp,reason}:
  2026-05-07T10:30,backend,postgres,destructive_op_blocked
  2026-05-07T11:45,frontend,figma,allowlist_violation
  2026-05-07T12:10,architect,redis,destructive_op_blocked

Retention (per spec §23.2)

AF_MCP_AUDIT_RETENTION_DAYS=30 (default) — older entries are pruned by a session-start sweep (NOT this skill).

Tie-Ins

• Spec: §9.10, §23 (MCP security)
• Hook: hooks/mcp-call-gate.cjs — sole writer of <security-dir>/mcp-audit.jsonl
• Resolver: hooks/lib/security-dir.cjs (shell mirror: scripts/plans/_lib.sh#security_dir) — where that log lives
• Script: scripts/security/sanitize-mcp-input.sh — sanitization run by the gate before logging
• Rule: rules/agent/mcp-security-policy.md — authoritative policy
• Rule: rules/agent/db-access-policy.md — DB-specific subset (architect/tdd-engineer only, read-only default, destructive ops blocked)
• Command: /aura-frog:mcp audit — primary user-facing entry to this skill
• Skill: Reuses scripts/json-to-toon.cjs to project the JSONL into context-friendly TOON