DEV Community

Cover image for Building 13 MCP Servers for US Government Data: From CVE Lookups to Disaster Alerts
Martin
Martin

Posted on

Building 13 MCP Servers for US Government Data: From CVE Lookups to Disaster Alerts

#mcp #ai #typescript #opensource

How I built and deployed 13 production MCP servers that give AI agents real-time access to US government data — cybersecurity vulnerabilities, safety recalls, natural disasters, and more.

If you've used Claude Desktop, Cursor, or any AI assistant that supports tool calling, you've probably wondered: what if my AI could pull live data from real sources instead of relying on training data?

That's exactly what MCP (Model Context Protocol) enables. It's an open standard by Anthropic that lets AI assistants call external tools — databases, APIs, services — through a standardized interface.

I built 13 production MCP servers that connect AI agents to free US government APIs. All open source, all deployed, all published to the Official MCP Registry.

GitHub: github.com/martc03/gov-mcp-servers

What I Built

The Servers

Category Server Tools Data Sources
Cybersecurity cybersecurity-vuln-mcp 7 NIST NVD 2.0, CISA KEV, EPSS, MITRE ATT&CK
Safety us-safety-recalls-mcp 4 NHTSA vehicle recalls, FDA food/drug recalls
Disasters natural-disaster-intel-mcp 4 FEMA, NOAA weather, USGS earthquakes
Finance federal-financial-intel-mcp 4 SEC EDGAR, CFPB complaints, BLS employment
Contracts gov-contracts-mcp 4 SAM.gov, USAspending
Legal court-records-mcp 4 Federal courts, PACER
Health public-health-mcp 4 NIH clinical trials, FDA adverse events
Business business-entity-mcp 4 SEC company search, SBA
Travel immigration-travel-mcp 3 Visa bulletins, border wait times
Environment environmental-compliance-mcp 3 EPA air quality, HUD foreclosures
Regulations regulatory-monitor-mcp 4 Federal Register, regulations.gov
Grants grant-finder-mcp 4 Grants.gov, USAspending
Intel competitive-intel-mcp 4 SEC filings, patents, trade data

Plus a REST API gateway with 45 endpoints across all 13 categories.

The Key Insight

Every one of these data sources is a free government API. No API keys required for most of them. Zero data cost. The US government publishes an enormous amount of structured data through public APIs — most developers just don't know they exist.

Architecture

Every server follows the same pattern:

User → MCP Client (Claude/Cursor) → Streamable HTTP → MCP Server → Government API
Enter fullscreen mode Exit fullscreen mode

Tech Stack

  • TypeScript — Type safety for API response parsing
  • @modelcontextprotocol/sdk — Official MCP SDK for tool definitions
  • Express — HTTP server for Streamable HTTP transport
  • Apify — Hosting platform (standby mode = long-running HTTP server)
  • Docker — Consistent deployment across all 13 servers

The Server Pattern

Every server follows an identical structure:

import { Actor } from "apify";
import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
import express from "express";

await Actor.init();

// Health check gate — exit cleanly for non-standby runs
if (process.env.APIFY_META_ORIGIN !== "STANDBY") {
    await Actor.pushData({ status: "healthy" });
    await Actor.exit("Health check passed");
}

// Create MCP server with tools
const mcp = new McpServer({
    name: "my-server",
    version: "1.0.0",
});

// Define tools
mcp.tool("search_recalls", { query: z.string() }, async ({ query }) => {
    const data = await fetch(`https://api.fda.gov/food/recall.json?search=${query}`);
    return { content: [{ type: "text", text: JSON.stringify(data) }] };
});

// Wire up Express + MCP transport
const app = express();
app.post("/mcp", async (req, res) => {
    const transport = new StreamableHTTPServerTransport("/mcp", res);
    await mcp.connect(transport);
    await transport.handleRequest(req, res);
});

const port = process.env.APIFY_ACTOR_STANDBY_PORT || 3000;
app.listen(port);
Enter fullscreen mode Exit fullscreen mode

This pattern repeats across all 13 servers. The only thing that changes is the tool definitions and the government APIs they call.

The Cybersecurity Server: A Deep Dive

The cybersecurity server is the most complex one, querying 4 data sources in parallel using Promise.allSettled:

Data Sources

  1. NIST NVD 2.0 — The National Vulnerability Database. Full CVE details, CVSS scores (v2, v3, v4), descriptions, references.

  2. CISA KEV — The Known Exploited Vulnerabilities catalog. If a CVE is in KEV, it's been actively exploited in the wild and federal agencies have mandatory remediation deadlines.

  3. FIRST.org EPSS — Exploitation Prediction Scoring System. A probability score (0-1) predicting the likelihood of exploitation in the next 30 days.

  4. MITRE ATT&CK — Maps CVEs to adversary techniques and tactics. I pre-computed a mapping of 172 CVEs to 42 techniques across 12 tactics from the STIX bundle.

Enriched Lookup

When you query vuln_lookup_cve with a CVE ID like CVE-2021-44228 (Log4Shell), all 4 sources are queried simultaneously:

CVSS: 10.0 CRITICAL
EPSS: 0.944 (94.4% exploitation probability, 99th percentile)
KEV: YES — Remediation due 2021-12-24
ATT&CK: T1190 (Exploit Public-Facing Application),
         T1203 (Exploitation for Client Execution),
         T1595.002 (Vulnerability Scanning)
Enter fullscreen mode Exit fullscreen mode

This kind of enriched view would normally require manually checking 4 different websites. Now an AI agent gets it in one tool call.

Lessons Learned

1. The Dockerfile Gotcha

The biggest debugging headache was a stale dist/ directory. Our Dockerfile uses a multi-stage build:

FROM node:18 AS builder
COPY . ./
RUN npm run build

FROM node:18
COPY --from=builder /home/myuser/dist ./dist
COPY . ./  # This overwrites builder's dist with local stale dist!
Enter fullscreen mode Exit fullscreen mode

That final COPY . ./ copies everything from the local build context — including a potentially stale dist/ folder — overwriting the freshly built one from the builder stage.

Fix: Added .dockerignore excluding dist/ and always running npm run build locally before pushing.

2. Apify Health Check Detection

Apify's platform runs periodic health checks on actors. For MCP servers in standby mode (long-running HTTP servers), these health checks would start the actor as a normal run, causing it to hang waiting for HTTP requests that never come — resulting in timeouts.

Fix: Check APIFY_META_ORIGIN:

  • "STANDBY" = normal MCP server mode
  • "API" = health check run → push data and exit immediately

3. Government APIs Are Better Than You Think

Many of these APIs are surprisingly well-designed:

  • NVD 2.0 has clean JSON responses with comprehensive CVE data
  • FEMA's API supports geospatial queries
  • FDA's openFDA API supports complex Elasticsearch-style queries
  • BLS has structured time-series data going back decades

The documentation can be sparse, but the APIs themselves are solid.

How to Use These Servers

Claude Desktop

Add to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "cybersecurity": {
      "url": "https://cybersecurity-vuln-mcp.apify.actor/mcp"
    },
    "safety-recalls": {
      "url": "https://us-safety-recalls-mcp.apify.actor/mcp"
    },
    "disasters": {
      "url": "https://natural-disaster-intel-mcp.apify.actor/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Any MCP Client

All 13 servers are published to the Official MCP Registry. Search for io.github.martc03 to find them all.

If you need a custom MCP server for your business, I build these professionally. Check out mcpdev.netlify.app or reach out at codee.mcpdev@gmail.com.

All 13 servers are open source: github.com/martc03/gov-mcp-servers

Top comments (0)