An MCP (Model Context Protocol) server for managing OWASP ModSecurity CRS via Docker. Gives AI assistants like Claude direct access to WAF monitoring, analysis, and configuration through a structured drill-down pipeline.

Built for Claude Code but works with any MCP-compatible client.

LLM proxy services (LiteLLM, OpenRouter, etc.) sit behind WAFs that generate massive amounts of false positives — prompts contain code, SQL, HTML, shell commands that trigger every content-inspection rule in the book. Managing these WAFs requires constant monitoring, tuning exclusions, and investigating events.

This MCP server lets an AI assistant do that work directly:

1. Overview — see total events, unique IPs, active rules at a glance
2. Drill down — filter events by IP or rule, inspect matched data
3. Act — disable rules, whitelist IPs, change engine mode — all without leaving the conversation

since — All analysis tools accept a since parameter to control the time window. Default is "24h". Supports Docker duration syntax: "1h", "24h", "7d", "30m". Days are automatically converted to hours (Docker's --since doesn't support the d suffix natively).

waf_overview(since: "7d")        # last 7 days
waf_events_by_ip(ip: "1.2.3.4", since: "1h")  # last hour

verbose — waf_events_by_ip, waf_events_by_rule, and waf_event_detail accept verbose: true. By default, matchedData and requestBody are truncated to keep responses within context limits:

• Docker with a running owasp/modsecurity-crs container
• Docker Compose managing the ModSecurity container
• Node.js 18+
• ModSecurity configured with JSON Serial audit log (SecAuditLogFormat JSON)

git clone https://github.com/KratosUAE/waf_mcp.git
cd waf_mcp
npm install
npm run build

claude mcp add --transport stdio --scope user \
  -e WAF_COMPOSE_DIR=/path/to/your/compose/dir \
  -e WAF_DOMAIN=https://your-domain.com \
  waf -- node /path/to/waf_mcp/dist/index.js

Or manually add to ~/.claude.json:

{
  "mcpServers": {
    "waf": {
      "type": "stdio",
      "command": "node",
      "args": ["/path/to/waf_mcp/dist/index.js"],
      "env": {
        "WAF_COMPOSE_DIR": "/path/to/your/compose/dir",
        "WAF_DOMAIN": "https://your-domain.com"
      }
    }
  }
}

The server expects a ModSecurity container managed by Docker Compose. Example service definition:

modsecurity:
  image: owasp/modsecurity-crs:nginx-alpine
  environment:
    - BACKEND=http://your-app:8080
    - MODSEC_RULE_ENGINE=DetectionOnly
    - MODSEC_AUDIT_LOG=/dev/stderr
    - MODSEC_AUDIT_LOG_FORMAT=JSON
    - MODSEC_AUDIT_LOG_TYPE=Serial
    - MODSEC_AUDIT_ENGINE=RelevantOnly
    - MODSEC_REQ_BODY_ACCESS=On
    - MODSEC_REQ_BODY_LIMIT=52428800
    - MODSEC_RESP_BODY_ACCESS=Off
    - PARANOIA=1
    - ANOMALY_INBOUND=5
  volumes:
    - ./modsecurity/REQUEST-900-EXCLUSION-RULES-BEFORE-CRS.conf:/etc/modsecurity.d/owasp-crs/rules/REQUEST-900-EXCLUSION-RULES-BEFORE-CRS.conf:ro

Key settings:

• MODSEC_AUDIT_LOG=/dev/stderr — sends audit log to Docker logs (required for the MCP server to read events)
• MODSEC_AUDIT_LOG_FORMAT=JSON — JSON format for structured parsing
• Exclusions file mount — allows hot-reload of rule exclusions via nginx -s reload

LLM API endpoints receive prompts containing code, SQL, HTML, and shell commands — all legitimate content that triggers WAF rules. Create an exclusions file to disable content-inspection rules on API paths:

# modsecurity/REQUEST-900-EXCLUSION-RULES-BEFORE-CRS.conf
SecRule REQUEST_URI "@rx ^(/v1/)?(chat/completions|completions|embeddings|responses|messages)|^/anthropic/" \
    "id:1000,phase:1,nolog,pass,\
    ctl:ruleRemoveById=921000-944999"

This disables rules 921000–944999 (all content-inspection categories: SQLi, XSS, RCE, LFI, RFI, etc.) on LLM API endpoints while keeping protocol enforcement, scanner detection, DoS protection, and IP reputation checks active.

Typical workflow in Claude Code:

You: "Check the WAF — anything suspicious?"

Claude: [calls waf_overview]
  → 332 events, 4 unique IPs, 12 rules triggered

Claude: [calls waf_top_ips]
  → 135.237.83.23 (Washington, US, Microsoft) — 320 events

Claude: [calls waf_events_by_ip, ip: "135.237.83.23", count: 5]
  → All POST /chat/completions, HTTP 200, rules: 942360, 932100...

Claude: [calls waf_event_detail, index: 42]
  → User-Agent: OpenAI/JS 6.26.0, body contains tool descriptions
  → Rule 942360 matched "update" in cron action descriptions

Claude: "This is your OpenClaw bot — all false positives.
         Want me to whitelist this IP?"

You: "Yes"

Claude: [calls waf_allow_ip, ip: "135.237.83.23"]
  → Done. IP whitelisted.

Investigating older events:

You: "Check IP 185.206.249.230 — it was flagged yesterday"

Claude: [calls waf_events_by_ip, ip: "185.206.249.230", since: "7d"]
  → 2 events from Apr 7, GET /v1/skills, HTTP 401, no rules triggered
  → Apple Private Relay IP (Singapore), just unauthorized API probes

npm run build        # Compile TypeScript
npm test             # Run tests (43 tests)
npm run test:watch   # Watch mode
WAF_DEBUG=1 npm start  # Run with debug logging

src/
├── index.ts           # MCP server setup, tool registration
├── waf-manager.ts     # Core service: Docker exec, log parsing, config management
├── types.ts           # TypeScript interfaces
├── config.ts          # Environment-based configuration
├── logger.ts          # stderr-only logger (stdout reserved for MCP protocol)
└── tools/
    ├── overview.ts        # L0: dashboard
    ├── top-ips.ts         # L1: IP aggregation
    ├── top-rules.ts       # L1: rule aggregation
    ├── fp-candidates.ts   # L1: false positive detection
    ├── events-by-ip.ts    # L2: drill-down by IP
    ├── events-by-rule.ts  # L2: drill-down by rule
    ├── event-detail.ts    # L3: full event inspection
    ├── status.ts          # Container status
    ├── set-engine.ts      # Engine mode control
    ├── set-paranoia.ts    # Paranoia level control
    ├── disable-rule.ts    # Rule management
    ├── enable-rule.ts     # Rule management
    ├── allow-ip.ts        # IP whitelist
    ├── deny-ip.ts         # IP whitelist
    ├── test.ts            # WAF test suite
    └── utils.ts           # Shared utilities

Events are parsed from Docker logs and cached for 30 seconds. Rapid drill-down calls (overview → top IPs → events by IP → event detail) hit the cache instead of re-parsing. The cache is invalidated when the since parameter changes.

MIT