🛡️ CVE MCP Server

AI-powered security intelligence at your fingertips — 28 tools + a one-call triage_cve orchestrator, 24 data sources, one protocol.

A production-grade Model Context Protocol (MCP) server that turns Claude into a full-spectrum security analyst. Instead of juggling 15+ browser tabs across NVD, EPSS, CISA KEV, Shodan, VirusTotal, and GreyNoise, ask Claude one question and get correlated intelligence in seconds. Built with Python, FastMCP, httpx, aiosqlite, Pydantic v2, and defusedxml.

The problem: Triaging a single CVE means querying NVD for CVSS scores, EPSS for exploitation probability, CISA KEV for active exploitation status, GitHub for patches, and VirusTotal for malware associations — then mentally correlating everything. For 50 CVEs, that's an entire day lost.

The solution: CVE MCP Server gives Claude direct access to 28 security tools across 24 APIs — fronted by the triage_cve one-call orchestrator. Ask "Should we patch CVE-2024-3400?" and Claude fans out to every relevant source in parallel, calculates a composite risk score (with a CISA KEV hard override), and delivers a prioritized recommendation with evidence.

🌍 GARS-2026 — Global Agentic AI Readiness Survey

I'm running a global academic study measuring how ready security professionals, developers, and enterprise teams actually are for agentic AI — MCP servers, tool calling, governance, and human-in-the-loop workflows.

If you use this repo, your response would be a genuinely valuable data point.

📋 Take the survey (10 min): Survey

• 60 questions · Anonymous · Supervised by SRH Berlin
• You get 50 Casky Tokens for early access to casky.ai
• Results published open access under CC-BY 4.0

📑 Table of contents

• Architecture
• Tool catalog
• Installation
• API keys setup
• Configuration
• Quick start
• Usage examples
• Risk score explained
• Data sources
• Running tests
• Architecture deep dive
• Security and privacy
• Troubleshooting
• Roadmap and known limitations
• Contributing
• License

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────────┐
│                        Claude Desktop / Claude Code                 │
│                         (MCP Client via stdio)                      │
└──────────────────────────────┬──────────────────────────────────────┘
                               │ Model Context Protocol (stdio)
                               ▼
┌─────────────────────────────────────────────────────────────────────┐
│                        CVE MCP Server (Python)                      │
│  ┌─────────────┐  ┌──────────────┐  ┌───────────────┐              │
│  │  27 MCP      │  │  Composite   │  │  SQLite Cache │              │
│  │  Tools       │  │  Risk Engine │  │  + Audit Log  │              │
│  └──────┬──────┘  └──────┬───────┘  └───────┬───────┘              │
│         │                │                   │                      │
│  ┌──────┴────────────────┴───────────────────┴──────┐               │
│  │              Async HTTP Client (httpx)            │               │
│  │         Rate Limiter · Response Cache             │               │
│  └──────────────────────┬───────────────────────────┘               │
└─────────────────────────┼───────────────────────────────────────────┘
                          │ HTTPS (outbound only)
          ┌───────────────┼───────────────────────────┐
          ▼               ▼                           ▼
┌──────────────┐ ┌──────────────┐            ┌──────────────┐
│ VULNERABILITY│ │   NETWORK    │            │   THREAT     │
│ INTELLIGENCE │ │ INTELLIGENCE │            │ INTELLIGENCE │
├──────────────┤ ├──────────────┤            ├──────────────┤
│ NVD API 2.0  │ │ AbuseIPDB    │            │ VirusTotal   │
│ EPSS / FIRST │ │ GreyNoise v3 │            │ MalwareBazaar│
│ CISA KEV     │ │ Shodan       │            │ ThreatFox    │
│ OSV.dev      │ │ CIRCL PDNS   │            │ Ransomwhere  │
│ GitHub GHSA  │ │              │            │ AlienVault   │
│ MITRE ATT&CK │ │              │            │ URLScan.io   │
└──────────────┘ └──────────────┘            └──────────────┘

All traffic is outbound HTTPS only — no inbound ports are opened. API keys are loaded from environment variables and never logged. Private/internal IP addresses are blocked from all lookup tools.

🔍 Tool catalog (28 tools)

⭐ Orchestration (v0.2.0) — start here

Also exposed via MCP primitives — Resources: kev://catalog, epss://scores/{cve_id}, manifest://tool-hash (SHA-256 over the registered tool surface, for tamper detection). Prompts: patch_decision, compare_and_prioritize, dependency_triage.

Core Vulnerability Intelligence (8 tools)

Exploit & Attack Intelligence (4 tools)

Phase 3: Advanced Risk & Reporting (4 tools)

Network Intelligence (4 tools)

Threat Intelligence (4 tools)

DevSecOps (3 tools)

📊 Risk score explained

The calculate_risk_score tool produces a composite risk score from 0 to 100 by weighting four independent signals.

The formula

Risk Score = (CVSS × 0.20) + (EPSS × 0.35) + (KEV × 0.30) + (PoC × 0.15)

Boost multipliers

• KEV + active PoC → ×1.15
• CVSS ≥ 9.0 + EPSS > 0.7 → ×1.10
• Published < 7 days ago → ×1.05

Score is capped at 100.

Risk Scoring — v1 (2026-06)

The numeric scorer is scoring_version 1.0 (surfaced in triage_cve, calculate_risk_score, and health_check). The linear weighted sum above is the v1 default for the numeric score, with one hard override:

• CISA KEV hard override: a KEV-listed CVE is confirmed exploited in the wild, the single strongest exploitation signal. Its label can never be lower than CRITICAL and its score is clamped to ≥ 76, regardless of CVSS/EPSS. (A KEV CVE with a low CVSS and low EPSS still returns CRITICAL / 76.)
• CVSS is treated as a severity signal, not an exploitation-likelihood signal (per Allodi & Massacci 2014); EPSS and KEV carry the exploitation signal.
• An experimental SSVC v2 gated decision (CISA Deployer model → Act / Attend / Track* / Track) is available via triage_cve(depth="deep") as a qualitative, explainable alternative to the 0–100 number.

Why these weights?

EPSS gets the highest weight (35%) because it's the single best predictor of actual exploitation — far better than CVSS alone. A CVSS 10.0 with EPSS 0.01 is theoretically dangerous but practically unlikely. KEV at 30% is ground truth: confirmed exploitation, not a prediction. CVSS at 20% captures severity context for new CVEs with insufficient EPSS data. PoC at 15% reflects that public exploits dramatically accelerate real-world attacks.

🆕 What's new in v0.2.0

• triage_cve orchestrator — one tool call that fans out NVD + EPSS + CISA KEV (+ public PoC discovery for depth != "quick") concurrently, computes the composite risk score, and returns a clean report. depth is quick / standard (default) / deep; deep additionally emits the SSVC v2 gated decision.
• New upstream sources — VulnCheck NVD++ (a transparent NVD fallback used automatically inside triage_cve when NIST NVD is unreachable/throttled), CIRCL hashlookup, and the HIBP Pwned Passwords range API.
• KEV hard-override scoring + scoring_version — KEV-listed CVEs are always CRITICAL (score ≥ 76); the scoring version is reported in triage_cve and health_check.
• HTTP transport — set MCP_TRANSPORT=http to serve streamable-HTTP on HOST:PORT (default 0.0.0.0:8000, stateless) instead of stdio. Ships with a Dockerfile.
• Resources & prompts — resources kev://catalog, epss://scores/{cve_id}, and manifest://tool-hash (SHA-256 over the registered tool surface); prompts patch_decision, compare_and_prioritize, and dependency_triage.
• Security posture — the server never registers a sampling handler / never issues sampling/createMessage (Unit 42 MCP-sampling attack vector); new outbound paths are scheme/host-allowlisted.

🌐 Data sources

🏛️ Architecture deep dive

File structure

src/cve_mcp/
├── server.py              # FastMCP server — all 27 @mcp.tool() definitions
├── config.py              # Environment config and API base URLs
├── models.py              # Pydantic models (CVERecord, KEVEntry, EPSSScore, ...)
├── audit.py               # Rotating audit log (50MB, 5 backups)
├── api/
│   ├── nvd_client.py      # NVD REST API v2.0
│   ├── osv_client.py      # OSV.dev package vulnerability API
│   ├── epss_client.py     # FIRST EPSS API
│   ├── kev_client.py      # CISA KEV catalog
│   ├── ip_intel.py        # AbuseIPDB + GreyNoise
│   ├── domain_intel.py    # crt.sh + CIRCL passive DNS
│   ├── shodan_client.py   # Shodan host intelligence
│   ├── hash_intel.py      # MalwareBazaar + VirusTotal
│   ├── url_safety.py      # URLScan.io
│   ├── malware_intel.py   # ThreatFox IOC lookup
│   ├── ransomware_intel.py# Ransomwhere Bitcoin address lookup
│   ├── exploit_intel.py   # GitHub PoC/exploit search
│   ├── vendor_advisory.py # MSRC + Red Hat + Ubuntu advisories
│   ├── attack_mapping.py  # MITRE ATT&CK STIX mapping
│   ├── cve_timeline.py    # CVE event timeline builder
│   ├── dependency_scan.py # OSV-based dependency scanning
│   ├── poc_checker.py     # GitHub + Exploit-DB + Nuclei PoC search
│   ├── report_generator.py# Vuln report + CVE comparison matrix
│   └── rate_limiter.py    # Token bucket rate limiter for NVD
├── cache/
│   └── sqlite_cache.py    # Async SQLite cache with per-key TTL
└── utils/
    ├── validators.py       # CVE ID normalization, IP/hash validation
    └── risk_scorer.py      # Composite risk score computation

Caching strategy

Audit log

Every tool invocation is logged to ~/.cve-mcp/audit.log:

{
  "timestamp": "2026-04-14T10:23:45.123Z",
  "tool": "lookup_cve",
  "parameters": {"cve_id": "CVE-2024-3400"},
  "duration_ms": 342,
  "cache_hit": false,
  "status": "ok"
}

API keys and response payloads are never written to audit logs.

🔐 Security and privacy

What data leaves your machine

• Outbound HTTPS only — no inbound ports opened, no telemetry
• CVE IDs, IPs, hashes, domains, and package names are sent to respective APIs for lookup
• API responses are cached locally in SQLite — cached data stays on your machine

Private IP blocking

All network intelligence tools block private and reserved IP ranges before any external API call:

• 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16 (RFC 1918)
• 127.0.0.0/8 (loopback), 169.254.0.0/16 (link-local)
• ::1, fc00::/7 (IPv6 private)

API key protection

• Keys loaded from environment variables only — never hardcoded
• .env is gitignored
• Keys never logged, cached, or included in audit entries

XML safety

defusedxml is used for all XML parsing to prevent XML bomb attacks (billion laughs, XXE injection).

🤝 Contributing

Contributions are welcome.

Adding a new tool

1. Add the tool function in server.py with the @mcp.tool() decorator
2. Add input validation in utils/validators.py
3. Implement the API client in api/
4. Add tests in tests/
5. Update this README

@mcp.tool()
async def my_new_tool(param: str, ctx: Context = None) -> str:
    """
    One-line description for Claude to know when to use this tool.

    Args:
        param: Description of the parameter
    """
    app = _get_app(ctx)
    # validate → cache check → API call → cache write → audit → return

Testing requirements

• All new tools must have at least one offline test with mocked responses
• Risk score changes must include formula verification test cases
• Network tools must include a test verifying private IP blocking
• All tests must pass: pytest tests/ -v

📄 License

MIT License — see LICENSE for details.

Copyright (c) 2025-2026 Mahipal Jangra (mukul975)

📦 Installation

Prerequisites

• Python 3.10+ (3.11 or 3.12 recommended)
• pip or uv package manager
• Git for cloning the repository
• A terminal with access to environment variables

Step-by-step setup

# 1. Clone the repository
git clone https://github.com/mukul975/cve-mcp-server.git
cd cve-mcp-server

# 2. Create and activate a virtual environment
python -m venv venv

# macOS / Linux:
source venv/bin/activate

# Windows (PowerShell):
.\venv\Scripts\Activate.ps1

# Windows (CMD):
venv\Scripts\activate.bat

# 3. Install dependencies
pip install -e .

# 4. Copy and configure environment variables
cp .env.example .env
# Edit .env with your API keys (see API Keys Setup section below)

# 5. Verify the server starts
python -m cve_mcp.server

Using uv (faster alternative)

git clone https://github.com/mukul975/cve-mcp-server.git
cd cve-mcp-server
uv venv
source .venv/bin/activate  # or .venv\Scripts\activate on Windows
uv pip install -e .
cp .env.example .env

With test dependencies

pip install -e ".[test]"

🔑 API keys setup

API keys are organized by priority — get the Tier 1 keys first for maximum coverage with free tools, then progressively add Tier 2 and Tier 3 as needed.

Tier 1: High priority (free, instant access, maximum coverage)

Tier 2: Recommended (free accounts, significant value)

Tier 3: Optional (extended intelligence)

⚡ Zero-key start: Eight tools work without any API key — EPSS, CISA KEV, OSV.dev, MITRE ATT&CK, CWE lookups, CVSS parsing, Ransomwhere, and NVD (at reduced rate). You can start using the server immediately and add keys progressively.

⚙️ Configuration

Environment variables (.env.example)

# NVD API key — free at https://nvd.nist.gov/developers/request-an-api-key
# Without key: 5 req/30s  |  With key: 50 req/30s
NVD_API_KEY=

# GitHub token — increases rate limit from 60/hr to 5000/hr (no scopes needed)
GITHUB_TOKEN=

# Threat intelligence keys (all optional — tools degrade gracefully without them)
ABUSEIPDB_KEY=       # https://www.abuseipdb.com/account/api
VIRUSTOTAL_KEY=      # https://www.virustotal.com/gui/join-us
URLSCAN_KEY=         # https://urlscan.io/user/signup
SHODAN_KEY=          # https://account.shodan.io/register

# GreyNoise — uses /v3/ip/{ip} endpoint (NOT the deprecated /v3/community)
GREYNOISE_API_KEY=   # https://viz.greynoise.io/signup

# CIRCL Passive DNS — requires partner registration
CIRCL_PDNS_USER=
CIRCL_PDNS_PASS=

# Optional overrides
CACHE_DB_PATH=       # defaults to ~/.cve-mcp/cache.db
AUDIT_LOG_PATH=      # defaults to ~/.cve-mcp/audit.log
REQUEST_TIMEOUT=30   # HTTP timeout in seconds
MAX_RETRIES=3        # retries on transient errors

Claude Desktop configuration

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "cve-mcp": {
      "command": "python",
      "args": ["-m", "cve_mcp.server"],
      "cwd": "/absolute/path/to/cve-mcp-server",
      "env": {
        "NVD_API_KEY": "your-key-here",
        "GITHUB_TOKEN": "ghp_xxxxxxxxxxxxxxxxxxxx",
        "ABUSEIPDB_KEY": "your-abuseipdb-key",
        "GREYNOISE_API_KEY": "your-greynoise-key",
        "SHODAN_KEY": "your-shodan-key"
      }
    }
  }
}

⚠️ Important: Always use absolute paths. Fully quit Claude Desktop (Cmd+Q / Alt+F4) after changing the config — reloading is not enough.

Claude Code configuration

# Basic setup
claude mcp add cve-mcp -- python -m cve_mcp.server

# With environment variables (repeat -e for each key)
claude mcp add cve-mcp -e NVD_API_KEY=your_key -e VULNCHECK_TOKEN=your_token -- python -m cve_mcp.server
# Or just run from the project directory — python-dotenv auto-loads .env on startup

# Verify it's connected
claude mcp list

🚀 Quick start

Step 1: Install (2 minutes)

git clone https://github.com/mukul975/cve-mcp-server.git
cd cve-mcp-server
python -m venv venv && source venv/bin/activate
pip install -e .

Step 2: Test with free tools first

No .env file needed. Add the server to Claude Desktop or Claude Code and try:

"What is CVE-2021-44228? Is it actively exploited?"

Claude will use lookup_cve (NVD), get_epss_score (EPSS), and check_kev (CISA KEV) — all free, no keys required.

"Scan these Python packages for vulnerabilities: requests 2.28.0, flask 2.2.0, django 3.2.0"

Step 3: Add your first key for 10× performance

echo 'NVD_API_KEY=your-key-here' > .env

Request a free NVD key at nvd.nist.gov — instant via email, increases rate limit from 5 to 50 requests per 30 seconds.

Step 4: Full power mode

Once you've added Tier 1 and Tier 2 keys:

"Calculate the risk score for CVE-2024-3400 and tell me if we should patch immediately."

💬 Usage examples

Scenario 1: "Should we patch Log4Shell immediately?"

You: Analyze if CVE-2021-44228 (Log4Shell) needs immediate patching. Give me the risk score and your recommendation.

Claude orchestrates multiple tools behind the scenes:

→ lookup_cve("CVE-2021-44228")
  CVSS 3.1: 10.0 (CRITICAL) | AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H

→ get_epss_score("CVE-2021-44228")
  EPSS: 0.97531 (97.5%) | Percentile: 99.99%

→ check_kev("CVE-2021-44228")
  ✅ IN CISA KEV — Added 2021-12-10 | Known ransomware use: Yes

→ check_poc_exists("CVE-2021-44228")
  ✅ WEAPONIZED — 300+ public PoC repositories found

→ calculate_risk_score("CVE-2021-44228")
  Risk Score: 97/100 (CRITICAL) | Urgency: PATCH IMMEDIATELY

Claude's response: "CVE-2021-44228 scores 97/100 (CRITICAL). Remote code execution in Apache Log4j, CVSS 10.0, 97.5% EPSS probability, confirmed active exploitation in CISA KEV including ransomware campaigns, 300+ public exploits. Patch immediately."

Scenario 2: "Scan my Python requirements.txt"

You: Scan these dependencies: requests==2.28.0, flask==2.2.0, django==3.2.0, pillow==9.0.0, cryptography==37.0.0

→ scan_dependencies(ecosystem="PyPI", packages={...})
  Found 12 vulnerabilities across 4 packages

Claude's response: "Found 12 vulnerabilities across your packages. Immediate action: Django 3.2.0 (3 critical CVEs including SQL injection), Pillow 9.0.0 (heap overflow), cryptography 37.0.0 (OpenSSL chain). Upgrade to: django>=4.2.16, pillow>=10.3.0, cryptography>=42.0.0."

Scenario 3: "Investigate a suspicious IP"

You: Investigate suspicious IP 185.220.101.34 — is it malicious?

→ check_ip_reputation("185.220.101.34")
  AbuseIPDB Confidence: 100% | Reports: 4,832 | ISP: Tor Exit Node

→ check_ip_noise("185.220.101.34")
  GreyNoise: MALICIOUS | Tags: [Tor Exit Node, Web Scanner]
  CVEs targeted: CVE-2024-3400, CVE-2023-46805

→ shodan_host_lookup("185.220.101.34")
  Open ports: 22, 80, 443, 9001 | OS: Linux

Claude's response: "185.220.101.34 is a known Tor exit node rated 100% malicious by AbuseIPDB with 4,832 abuse reports. GreyNoise confirms active scanning for CVE-2024-3400 (Palo Alto) and CVE-2023-46805 (Ivanti). Recommendation: Block at firewall."

Scenario 4: "Generate an executive security report"

You: Generate an executive security report for: CVE-2024-3400, CVE-2023-44487, CVE-2023-4966, CVE-2024-21887

→ compare_cves([...])
  Compiles: NVD, EPSS, KEV, PoC availability, MITRE ATT&CK mapping
  Generates: Ranked risk table, remediation timeline, executive summary

🧪 Running tests

# Run the full test suite
pytest tests/ -v

# Run specific test files
pytest tests/test_validators.py tests/test_risk_scorer.py -v

# Run with coverage
pytest tests/ -v --cov=src/cve_mcp --cov-report=term-missing

Test with the MCP Inspector

npx @modelcontextprotocol/inspector python -m cve_mcp.server

Opens at http://localhost:6274 — interactively test each tool, view input schemas, and inspect response formats.

What tests cover

• Unit tests: Risk score calculation, CVSS vector parsing, input validation
• Integration tests: Tool registration, parameter validation, error handling for missing keys
• Cache tests: SQLite cache writes, TTL expiration, cache hit/miss
• Security tests: Private IP blocking, XML bomb protection (defusedxml), input sanitization

🔧 Troubleshooting

Server won't start

# Ensure virtual environment is activated and package is installed
pip install -e .
python --version  # must be 3.10+

Claude Desktop doesn't show the hammer icon (🔨)

• Check for JSON syntax errors (no trailing commas) in your config
• Use absolute paths — relative paths silently fail
• Fully quit Claude Desktop (Cmd+Q / Alt+F4) and restart

NVD rate limited

# Add your free NVD API key to .env
NVD_API_KEY=your-key-here
# https://nvd.nist.gov/developers/request-an-api-key

The server queues excess requests automatically, but with a key you get 10× throughput.

GreyNoise 401 Unauthorized

# Verify your key works:
curl -H "key: YOUR_KEY" https://api.greynoise.io/v3/ip/8.8.8.8
# The server uses /v3/ip/{ip} — NOT the deprecated /v3/community endpoint

Windows encoding issues

$env:PYTHONUTF8 = "1"
$env:PYTHONIOENCODING = "utf-8"

🗺️ Roadmap and known limitations

What the server does NOT do

• No active scanning — intelligence/lookup only, does not probe your infrastructure
• No write operations — reads from external APIs only (except URLScan submissions)
• No CVSS v4.0 scoring — built-in calculator handles v3.1 only; NVD-provided v4.0 scores are displayed but not recalculated

Known API limitations

• NVD returns max 2,000 results per query
• EPSS scores for brand-new CVEs (< 24 hours old) may not exist yet
• CISA KEV updates on US business days only
• GreyNoise community tier: 50 queries/week
• VirusTotal free tier: 4 requests/minute
• CIRCL PDNS requires manual registration and approval
• Ransomwhere has a 90-day embargo on new addresses

Planned improvements

• CVSS v4.0 local calculator
• Webhook/alerting for KEV additions and EPSS score changes on a CVE watchlist
• STIX 2.1 export for SIEM integration
• Docker container with zero-install deployment
• Streamable HTTP transport (MCP SSE)
• Additional sources: Censys, SecurityTrails, VulnCheck