Cve MCP

The Problem

Vulnerability intelligence is scattered across multiple databases. NVD has CVE details. EPSS tells you exploitation probability. CISA KEV tracks actively exploited vulns. GitHub Advisory covers open source packages. OSV maps vulnerabilities to specific package versions. Shodan tracks internet exposure. Nuclei and Metasploit tell you if there's a working exploit. No single tool aggregates them, and none work with AI agents.

Traditional workflow:
  search NVD for CVE details          →  navigate a clunky web UI
  check EPSS for exploitation risk    →  separate API, separate format
  look up CISA KEV status             →  download a JSON feed manually
  search GitHub advisories            →  yet another interface
  query OSV for package impact        →  different API, different schema
  check Shodan for exposure           →  separate subscription
  look for Nuclei/MSF exploits        →  manual GitHub search
  map to ATT&CK techniques            →  separate MITRE lookup
  correlate everything                →  copy-paste into a spreadsheet
  ──────────────────────────────────
  Total: 30+ minutes per CVE, longer for bulk triage

cve-mcp gives your AI agent 41 tools via the Model Context Protocol. The agent queries 11 sources in parallel, correlates data, detects weaponization signals, calculates risk scores, and tells you exactly what matters.

With cve-mcp:
  You: "Prioritize these 10 CVEs by actual exploitation risk"

  Agent: → fetches CVSS scores from NVD
         → gets EPSS exploitation probability for each
         → checks CISA KEV for actively exploited
         → cross-references GitHub advisories for patches
         → "3 are critical: CVE-2024-3400 (EPSS 97%, in KEV),
            CVE-2023-44487 (HTTP/2 rapid reset, EPSS 96%),
            CVE-2021-44228 (Log4Shell, EPSS 97%, in KEV).
            Here are patches and affected versions..."

How It's Different

Existing tools give you raw data. cve-mcp gives your AI agent the ability to reason about vulnerabilities.

Quick Start

Option 1: npx (no install)

npx cve-mcp

Option 2: Clone

git clone https://github.com/badchars/cve-mcp.git
cd cve-mcp
bun install

Environment variables (all optional)

# Increases NVD rate limit from 5 to 50 requests per 30 seconds
export NVD_API_KEY=your-nvd-api-key

# Enables GitHub Advisory search (60 → 5000 requests/hour)
export GITHUB_TOKEN=ghp_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

# VulnCheck extended KEV, CPE, PURL search
export VULNCHECK_API_KEY=your-vulncheck-key

# Vulners vulnerability search across 200+ sources
export VULNERS_API_KEY=your-vulners-key

# AttackerKB community assessments
export ATTACKERKB_API_KEY=your-attackerkb-key

All are optional. The server works without them — some tools just have lower rate limits or reduced data.

Connect to your AI agent

# With npx
claude mcp add cve-mcp -- npx cve-mcp

# With local clone
claude mcp add cve-mcp -- bun run /path/to/cve-mcp/src/index.ts

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "cve-mcp": {
      "command": "npx",
      "args": ["cve-mcp"],
      "env": {
        "NVD_API_KEY": "optional-key",
        "GITHUB_TOKEN": "optional-token",
        "VULNCHECK_API_KEY": "optional-key",
        "VULNERS_API_KEY": "optional-key",
        "ATTACKERKB_API_KEY": "optional-key"
      }
    }
  }
}

Same JSON config format. Point the command to npx cve-mcp or your installation path.

Start querying

You: "What do you know about CVE-2024-3400?"

That's it. The agent handles the rest.

What The AI Can Do

Incident Response

You: "We got hit by CVE-2024-3400. Give me everything."

Agent: → cve_enrich {cveId: "CVE-2024-3400"}
       → NVD: PAN-OS command injection, CVSS 10.0
       → EPSS: 97.2% exploitation probability
       → KEV: Added 2024-04-12, due 2024-05-01
       → Shodan: 4 CPEs, internet-exposed
       → Nuclei: Detection template exists (critical)
       → Metasploit: Exploit module available (excellent rank)
       → "Critical. Weaponized — both Nuclei and MSF modules exist.
          PAN-OS GlobalProtect, versions < 10.2.9-h1.
          Patch immediately. CISA deadline: May 1."

Vulnerability Prioritization

You: "Prioritize these CVEs from our scan: CVE-2021-44228, CVE-2024-3400,
      CVE-2023-44487, CVE-2024-21762, CVE-2023-4966"

Agent: → cve_prioritize {cves: [...]}
       → Ranks by CVSS × EPSS × KEV multiplier
       → "#1: CVE-2024-3400 (risk: 19.44, CVSS 10.0, EPSS 97%, KEV)
          #2: CVE-2021-44228 (risk: 19.40, CVSS 10.0, EPSS 97%, KEV)
          #3: CVE-2023-4966 (risk: 19.10, CVSS 9.4, EPSS 97%, KEV)
          All 5 are in CISA KEV — patch all immediately."

Dependency Audit

You: "Check if lodash 4.17.20 and django 3.2.0 have known vulnerabilities"

Agent: → osv_batch {queries: [
          {package: "lodash", version: "4.17.20", ecosystem: "npm"},
          {package: "django", version: "3.2.0", ecosystem: "PyPI"}
       ]}
       → "lodash 4.17.20: 3 vulnerabilities (prototype pollution)
          django 3.2.0: 12 vulnerabilities (SQL injection, XSS)
          Upgrade lodash to 4.17.21+, django to 4.2+"

Threat Monitoring

You: "What are the most likely to be exploited CVEs right now?"

Agent: → cve_trending {limit: 10, minEpss: 0.9}
       → "Top 10 by exploitation probability:
          1. CVE-2024-3400 — PAN-OS (EPSS 97.2%, CVSS 10.0, KEV)
          2. CVE-2023-44487 — HTTP/2 Rapid Reset (EPSS 96.5%, CVSS 7.5, KEV)
          ..."

CVSS Deep Dive (v3.1 & v4.0)

You: "Break down this CVSS v4.0 vector: CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N"

Agent: → cvss_parse {vector: "CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N"}
       → "CVSS v4.0 — Score: 9.3 (CRITICAL)
          Network attack, low complexity, no attack requirements.
          No privileges, no user interaction.
          Vulnerable system: full C/I/A impact.
          Subsequent systems: no impact."

ATT&CK Mapping

You: "Map CVE-2024-3400 to MITRE ATT&CK techniques"

Agent: → cve_to_attack {cweIds: ["CWE-77"]}
       → "CWE-77 (Command Injection) maps to:
          T1059 — Command and Scripting Interpreter (Execution)
          T1190 — Exploit Public-Facing Application (Initial Access)"

Tools Reference (41 tools)

Data Sources

Risk Score Formula

Risk Score = CVSS Base Score × EPSS Score × KEV Multiplier × Exploit Multiplier

Where:
  CVSS Base Score     = 0-10 (severity from NVD)
  EPSS Score          = 0-1 (exploitation probability from FIRST)
  KEV Multiplier      = 2 if in CISA KEV, 1 otherwise
  Exploit Multiplier  = 1.5 if Nuclei template OR Metasploit module exists, 1 otherwise

This gives a practical risk score that balances severity (CVSS) with real-world exploitation likelihood (EPSS), known active exploitation (KEV), and weaponization signals (public exploit tooling).

Architecture

src/
├── index.ts                    Entry point + MCP stdio
├── types/
│   └── index.ts                ToolDef, ToolContext, ToolResult, API types
├── protocol/
│   ├── tools.ts                41 tool definitions (Zod schemas)
│   └── mcp-server.ts           MCP server + stdio transport
├── nvd/
│   ├── index.ts                NVD API v2 — search, get, recent
│   └── cpe.ts                  Product/CPE search
├── epss/
│   └── index.ts                EPSS — score, top
├── kev/
│   └── index.ts                KEV — check, search, recent (cached)
├── ghsa/
│   └── index.ts                GitHub Advisory — search, get
├── osv/
│   └── index.ts                OSV — query, get, batch
├── exploit/
│   └── index.ts                PoC search via GitHub repos
├── shodan/
│   └── index.ts                Shodan CVEDB + InternetDB (zero-auth)
├── vulncheck/
│   └── index.ts                VulnCheck KEV, CPE, PURL
├── vulners/
│   └── index.ts                Vulners search + lookup
├── nuclei/
│   └── index.ts                Nuclei template existence check
├── metasploit/
│   └── index.ts                MSF module check (cached metadata)
├── cpe/
│   └── index.ts                NVD CPE dictionary API
├── circl/
│   └── index.ts                CIRCL CVE enrichment
├── attackerkb/
│   └── index.ts                AttackerKB assessments
├── attack/
│   └── index.ts                CWE → MITRE ATT&CK mapping
├── cwe/
│   └── index.ts                CWE API (MITRE) + static fallback
├── cvss/
│   └── index.ts                CVSS v3.1 + v4.0 parser + calculator
├── meta/
│   ├── enrich.ts               Full CVE enrichment (8 sources parallel)
│   ├── prioritize.ts           Risk-based CVE ranking
│   ├── trending.ts             Trending CVEs by EPSS
│   ├── compare.ts              Side-by-side CVE comparison
│   └── sources.ts              11 source health checks
└── utils/
    ├── rate-limiter.ts          Queue-based rate limiter
    └── cache.ts                 TTL cache

Design decisions:

• Intelligence, not audit — Unlike cloud-audit-mcp and github-security-mcp, this is a data tool. No CheckResult, no findings accumulation. Each query is independent and stateless.
• Parallel enrichment — cve_enrich calls 8 sources via Promise.allSettled. If one source is down, the rest still return data.
• Weaponization detection — Checks Nuclei template and Metasploit module existence to flag CVEs with public exploit tooling.
• Shared rate limiter — All NVD modules share a single RateLimiter instance (6s between requests) to avoid 429 errors.
• KEV + MSF caching — KEV catalog (~1200 entries) and MSF module metadata (~15MB) loaded once, cached with 1-hour TTL.
• CWE dual-mode — MITRE CWE REST API for full detail (1000+ CWEs), with 40+ entry static fallback when API is unreachable.
• CVSS v3.1 + v4.0 — Auto-detects version from vector prefix. V4.0 uses MacroVector scoring approach.
• ATT&CK mapping — Static CWE-to-technique table (33 CWE entries → ATT&CK techniques). No API calls.
• 2 dependencies — @modelcontextprotocol/sdk and zod. Nothing else.

Limitations

• NVD API without NVD_API_KEY is limited to 5 requests per 30 seconds. Set the key for production use
• GitHub Advisory search without GITHUB_TOKEN is limited to 60 requests per hour
• VulnCheck, Vulners, and AttackerKB require API keys for full functionality
• Exploit search uses GitHub repository search which has its own rate limits
• MSF module metadata (~15MB) is loaded on first use — initial msf_check call is slower
• CVSS v4.0 scoring uses MacroVector approximation (exact specification is very complex)
• ATT&CK mapping covers 33 CWE entries — uncommon CWEs may not map
• macOS / Linux (Windows not tested)

Part of the MCP Security Suite