CMDB-CVE MCP Server

A Model Context Protocol (MCP) server for vulnerability intelligence automation. This system combines a Configuration Management Database (CMDB) with real-time CVE ingestion from NIST NVD and leverages LLM-driven reasoning for vulnerability correlation.

Overview

This MCP server enables an AI agent (e.g., Claude Desktop) to:

• Query a read-only CMDB (asset inventory)
• Fetch recent CVE data from NVD via the official REST API
• Perform LLM-driven, version-aware correlation between CVEs and assets
• Generate and send structured, explainable vulnerability reports via email

Data Flow

1. Load assets from CMDB
2. Fetch CVEs from NVD
3. Pre-process and analyze CVEs + assets
4. LLM performs version-aware correlation
5. Generate structured vulnerability findings
6. Send report via email

Architecture

cmdb-cve-mcp/
├── server.py          # MCP server (tool definitions + orchestration)
├── cmdb.py            # SQLite-based read-only CMDB with sample assets
├── nvd_client.py      # Client to query NVD CVE API
├── email_sender.py    # SMTP email utility
├── requirements.txt
└── README.md

Available MCP Tools

| Tool | Description | Example Prompt | | --------------------------- | --------------------------------------------------- | ------------------------------------------------------------ | | list_assets | View all CMDB assets (read-only) | "Show me all assets in my CMDB" | | get_assets_json | Retrieve assets as structured JSON for LLM analysis | "Get assets in JSON format" | | get_cves_json | Retrieve CVEs as structured JSON for LLM analysis | "Fetch CVEs in JSON format" | | fetch_latest_cves | Retrieve CVEs in human-readable format | "Show latest CVEs of last 3 days with CVSS > 9" | | send_vulnerability_report | Send report based on LLM-generated findings | "Send report to user@example.com" |

Guardrails

Security Guardrails

• Read-only CMDB – prevents asset modification
• Trusted data source – CVEs fetched only from official NVD API
• Restricted tool scope – no arbitrary command execution
• Secure configuration – credentials managed via environment variables

Operational Guardrails

• CVE window limits – capped query range (max 90 days)
• Rate-limit aware fetching – respects NVD API constraints
• Bounded outputs – prevents excessive token usage
• Guided LLM reasoning – enforces structured, version-aware decisions

Limitations

❗ No Native Version Parser

• Version extraction relies on LLM interpretation of strings
• May struggle with non-standard version formats and missing version data

❗ Scalability Issue

• All CVEs and assets may be passed to LLM for co-relation
• Can lead to increased latency, higher token usage and reduced efficiency at scale

❗ LLM Dependency

• Accuracy depends on prompt quality and model capability
• May cause hallucinations

❗ No Automated Validation

• No programmatic verification of LLM decisions
• Incorrect reasoning may pass through if not reviewed

❗ No Exploit Intelligence

• Does not consider active exploitation, public exploit availability and threat prioritization