nist-nvd-mcp-server

---

Overview

CVE and CPE data from the NIST National Vulnerability Database. Search and audit vulnerabilities by keyword, severity, CWE, or CISA KEV status, resolve products to CPE names, and track a CVE's revision history from any MCP client. Runs as a stdio process, a local Streamable HTTP server, or the public hosted endpoint above.

Tools

Tool	Description
nvd_search_cves	Search CVEs by keyword, severity, CWE, date range, or CISA KEV status.
nvd_get_cve	Fetch one or more CVEs by ID — full CVSS scores, CWE, CPE configs, KEV fields, and references.
nvd_search_cpes	Search the NVD CPE dictionary by product keyword or partial match string.
nvd_audit_cpe	Find all CVEs affecting a specific product version by CPE name or virtual match string.
nvd_get_cve_history	Retrieve the change history for a CVE — score revisions, status transitions, and reference additions.

Resources

Resource	Description
nvd://cve/{cveId}	Full CVE record by ID — same data as nvd_get_cve for a single ID, as a stable URI for injectable context.

All resource data is also reachable via tools.

Capability reference

nvd_search_cves _tool

• Full-text keyword search (AND-semantics across words), or exactPhrase: true for an exact-phrase match — requires keyword
• Filters: CVSS severity band (LOW/MEDIUM/HIGH/CRITICAL — CRITICAL requires severityVersion: "v3" or "v4"), CWE ID, CISA KEV status, noRejected (default true)
• Date filters: pubDays/lastModDays convenience shorthands (auto-clamped to 120 days, clamping reported in the enrichment) or explicit ISO 8601 ranges (120-day max span, both ends required); the two forms per axis are mutually exclusive
• Pagination via limit (up to 2000, default 20) and offset
• Always returns brief summaries with a truncated description; call nvd_get_cve for full detail

---

nvd_get_cve _tool

• Batch up to 100 CVE IDs per call
• Full mode (default): CVSS scores across v2.0/v3.0/v3.1/v4.0, CWE weaknesses, CPE configurations, CISA KEV fields, references
• brief: true returns trimmed rows (ID, status, top severity, KEV name, truncated description) — recommended for batches over 10
• includeReferences: false strips the references array; allLanguages: true renders every localized description instead of English-only
• missingIds enrichment field lists any requested IDs NVD didn't return
• Rendered text caps references at 15 per record, with a … N more trailer

---

nvd_search_cpes _tool

• Keyword search (e.g. "apache http server") or a partial CPEv2.3 pattern via cpeMatchString — at least one required
• Returns full CPE name, human-readable title, deprecation status, and superseding CPEs
• Pagination via limit (up to 10,000, default 20) and offset — a vendor-level keyword can match tens of thousands of entries, so page rather than narrowing further
• Use before nvd_audit_cpe to resolve the exact CPE name a product needs

---

nvd_audit_cpe _tool

• Two modes: exact cpeName (NVD auto-applies isVulnerable) or virtualMatchString with optional versionStart/versionEnd bounds (inclusive/exclusive)
• Client-side severityMin filter drops low-signal entries from the fetched page — it can only remove what limit already retrieved
• Returns full CVE records (CVSS scores, CWE, CPE configurations, KEV fields, references)
• Pagination via limit (up to 2000, default 20) and offset — page at a modest limit rather than raising it, since each result is a full record
• auditTarget enrichment field echoes the CPE identifier used, so callers can verify the correct product was queried

---

nvd_get_cve_history _tool

• Returns change events: CVSS revisions, status transitions, reference additions, CPE configuration updates
• order picks the anchor end — newest (default) reads most-recent-first, oldest reads NVD's native order
• Paginated via limit (up to 2000, default 20) and offset, counted from the end order anchors to
• The history endpoint is markedly slower without an API key — set NVD_API_KEY and raise NVD_REQUEST_TIMEOUT_MS

---

nvd://cve/{cveId} _resource

• Full CVE record as application/json — same data as nvd_get_cve for one ID, with references and English-only descriptions
• cveId must match CVE-YYYY-NNNNN; a well-formed but unknown ID throws cve_not_found

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

NVD-specific:

• Request pacer enforces NVD's 5 req/30s (no key) and 50 req/30s (with key) limits with automatic queuing, at a minimum inter-request gap derived from the window and limit
• Retry wraps the pacer rather than sitting inside it — every attempt takes its own turn in the queue, so retries count against the rate budget instead of bursting past it
• A 403's Retry-After holds the whole queue until NVD's window resets; keyless, a 403 fails fast and names NVD_API_KEY rather than spending the 5-request budget on retries that cannot outlast a 30-second window
• Deterministic rejections fail fast instead of consuming retries — NVD answers both a bad parameter and a refused API key with HTTP 404, separated only by a message header, so a refused key surfaces as a config fault naming NVD_API_KEY rather than as a malformed CVE ID
• HTML-response guard catches NVD rate-limit pages served as HTML instead of a 403

Agent-friendly output:

• An enrichment block on every response, carried on both structuredContent and the rendered text — total results, returned count, page offset, the filters actually applied, and any date-clamping events, so agents can reason about what was really queried
• missingIds in batch CVE lookups — a per-ID parity check instead of a silent partial result
• CPE echo in audit responses — cpeName or virtualMatchString reflected back so callers can verify the correct product was audited
• Empty-result notices that name the cause — an unmatched query, a severity threshold that emptied the page, an offset past the end of the result set, and a clean audit ("no known vulnerabilities") are told apart rather than all reading as errors or "nothing found"

Getting started

Public Hosted Instance

A public instance is available at https://nist-nvd.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "nist-nvd-mcp-server": {
      "type": "streamable-http",
      "url": "https://nist-nvd.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add the following to your MCP client configuration file.

{
  "mcpServers": {
    "nist-nvd-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/nist-nvd-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info",
        "NVD_API_KEY": "your-api-key"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "nist-nvd-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/nist-nvd-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info",
        "NVD_API_KEY": "your-api-key"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "nist-nvd-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT_TYPE=stdio",
        "-e", "NVD_API_KEY=your-api-key",
        "ghcr.io/cyanheads/nist-nvd-mcp-server:latest"
      ]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 NVD_API_KEY=... bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• Optional: NVD API key — free, raises rate limit from 5 req/30s to 50 req/30s.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/nist-nvd-mcp-server.git

2. Navigate into the directory:

cd nist-nvd-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# edit .env and set NVD_API_KEY if you have one

Configuration

Variable	Description	Default
NVD_API_KEY	NVD API key. Without it, rate limit is 5 req/30s; with it, 50 req/30s. Get one free at nvd.nist.gov/developers/request-an-api-key.	—
NVD_REQUEST_TIMEOUT_MS	Per-request timeout in milliseconds. The history endpoint is slow without an API key — raise to 60000 if using nvd_get_cve_history without a key.	10000
MCP_TRANSPORT_TYPE	Transport: stdio or http.	stdio
MCP_HTTP_PORT	Port for HTTP server.	3010
MCP_AUTH_MODE	Auth mode: none, jwt, or oauth.	none
MCP_LOG_LEVEL	Log level (RFC 5424).	info
LOGS_DIR	Directory for log files (Node.js only).	<project-root>/logs
OTEL_ENABLED	Enable OpenTelemetry instrumentation.	false

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http
  

• Run checks and tests:

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest test suite
  bun run lint:mcp   # Validate MCP definitions against spec
  

Docker

docker build -t nist-nvd-mcp-server .
docker run --rm -e NVD_API_KEY=your-key -p 3010:3010 nist-nvd-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/nist-nvd-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Directory	Purpose
src/index.ts	createApp() entry point — registers tools/resources and inits services.
src/config	Server-specific environment variable parsing and validation with Zod.
src/mcp-server/tools	Tool definitions (*.tool.ts).
src/mcp-server/resources	Resource definitions (*.resource.ts).
src/services/nvd-http	NVD HTTP client with rate-limit pacing and retry.
src/services/nvd-cve	CVE service — search, fetch-by-ID, CPE audit, change history, normalization.
src/services/nvd-cpe	CPE service — dictionary search and normalization.
src/services/nvd-source	Source service — resolves NVD contributor identifiers to their published names.
tests/	Unit and integration tests mirroring src/.

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools and resources via the barrels in src/mcp-server/*/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.