Endoflife MCP Server

endoflife.ai — MCP Server

Exposes endoflife.ai's lifecycle intelligence to AI agents over the Model Context Protocol (MCP, Streamable HTTP transport). A dependency-free Cloudflare Worker that wraps the public api.endoflife.ai/v1 endpoints and the site's published feeds — no data duplicated, every answer carries a source URL.

Current version: 1.4.1 (SERVER_INFO.version in src/index.js, server.json, package.json, the Dockerfile ARG VERSION, and the site's endoflife/.well-known/mcp/server-card.json — keep them in step; the registry and Glama read server.json).

1.4.1: description and tool text carry the current build-computed counts (700+ products, 15,000+ versions) for the registry and npm listings; no behaviour change.

1.4.0: check_sbom is a thin proxy to the API's POST /v1/sbom (one resolver for every surface; versions keep their release family, e.g. python 3.12, ubuntu 20.04). Above the tier cap it returns the scored subset plus limit_note and not_scored_count; scan_stack scores the first N the tier allows and lists the rest in not_checked. Until 1.4.0 check_sbom cut every version to its major number and a keyless SBOM with more than 5 matched components returned nothing.

Tools (all read-only)

Every tool advertises annotations (readOnlyHint, idempotentHint) and a permissive outputSchema; results are returned both as JSON text and as structuredContent. Lookup misses return "did you mean" slug suggestions (prefix / substring / edit-distance ≤ 2 against the live product list).

Resources (resources/list / resources/read): llms.txt, the EOS Edge Device feed, Exploited & Unpatchable, KEV by product, the edge change log.

Prompts (prompts/list / prompts/get): audit_stack, eol_calendar, edge_device_review.

Endpoints: POST / (JSON-RPC), GET / (info page), GET /health (liveness JSON), GET /.well-known/mcp/server-card.json (discovery card).

Usage telemetry

With the USAGE Analytics Engine binding (see wrangler.toml, dataset endoflife_mcp_usage) each method / tool call writes one data point: tool name, client user agent, ok/error, keyed/anon, latency in ms. No request bodies. Query via the Cloudflare Analytics Engine SQL API, e.g.

SELECT blob1 AS tool, count() AS calls, sum(_sample_interval) AS weighted
FROM endoflife_mcp_usage WHERE timestamp > NOW() - INTERVAL '7' DAY
GROUP BY tool ORDER BY calls DESC

The binding is optional; the code no-ops without it.

Deploy

Production deploys run from GitHub Actions, never a local wrangler:

gh workflow run mcp-server-deploy.yml

(.github/workflows/mcp-server-deploy.yml, pinned wrangler; needs the CLOUDFLARE_API_TOKEN / CLOUDFLARE_ACCOUNT_ID repository secrets.) Local wrangler dev still works for development; the API service binding and the USAGE dataset are declared for both the default and production environments.

Custom domain (mcp.endoflife.ai)

The route in wrangler.toml needs mcp.endoflife.ai to resolve through Cloudflare: an AAAA record, name mcp, IPv6 100::, proxied. Already in place.

Test before deploying

A Node harness exercises every method against the live API and feeds (the module runs unchanged in Node with env.API = { fetch }):

node scratch/mcp_test.mjs   # see the session scratchpad; 22 RPC calls, expects 0 failures

Offline unit tests (fetch mocked, no network): npm test in mcp-server/ (node --test test/).

Run as a container (Red Hat UBI)

server.mjs runs the same handler that serves mcp.endoflife.ai inside a plain Node.js process, and Dockerfile packages it on registry.access.redhat.com/ubi9/nodejs-22-minimal for cluster deploys (the form the OpenShift AI MCP catalog expects). No build step, no dependencies; the container talks to https://api.endoflife.ai over HTTPS and runs as the unprivileged UBI user (uid 1001).

docker build -t endoflife-mcp mcp-server/
docker run --rm -p 8080:8080 endoflife-mcp
curl -s localhost:8080/health

Endpoints are the Worker's own: POST / (Streamable HTTP JSON-RPC), GET /health (readiness), GET /.well-known/mcp/server-card.json. Set ENDOFLIFE_API_KEY (docker run -e ENDOFLIFE_API_KEY=...) and the server sends it as X-API-Key on every call that carries no key of its own. A self-hosted server needs it from November 16, 2026 at 00:00 UTC, when the API stops answering data requests without a key; only the hosted endpoint at mcp.endoflife.ai stays keyless. A free key is emailed at once from https://endoflife.ai/api#free-key (150 requests a day, no expiry; higher volume is on a paid key). Validated on RHEL 9 and 10 with Node 20 and 22 by .github/workflows/rhel-validation.yml (evidence in docs/company/redhat-validation-evidence.md).

Connect from an MCP client

Claude Desktop / Cursor / VS Code (mcpServers):

{
  "mcpServers": {
    "endoflife": { "command": "npx", "args": ["mcp-remote", "https://mcp.endoflife.ai"] }
  }
}

Clients with native remote-MCP support can use the URL directly:

{ "mcpServers": { "endoflife": { "url": "https://mcp.endoflife.ai" } } }

Test without a client

# tools/list
curl -s https://mcp.endoflife.ai -X POST -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq

# KEV exposure for a product
curl -s https://mcp.endoflife.ai -X POST -H 'Content-Type: application/json' \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"get_kev_exposure","arguments":{"product":"ivanti-connect-secure"}}}' | jq

# health + discovery card
curl -s https://mcp.endoflife.ai/health | jq
curl -s https://mcp.endoflife.ai/.well-known/mcp/server-card.json | jq

Auth & tiers

The read tools work with no key for individual use and evaluation through the hosted endpoint mcp.endoflife.ai (each agent gets 2,000 a day and 5 per call). Calling the API directly, or running this server yourself, works without a key only until November 16, 2026 at 00:00 UTC (anonymous tier: 50 requests a day, 3 components per SBOM or batch call); from then every data request needs a key (set ENDOFLIFE_API_KEY on a self-hosted server; free key at https://endoflife.ai/api#free-key). Agents that run inside a company's tooling, check inventories or call on a schedule are production use and belong on a paid key: Starter ($89/month, 10,000 requests a day, 25 per call) or Pro ($299/month, 100,000 requests a day under fair use, 50 per call), issued at checkout from https://endoflife.ai/api?utm_source=mcp#get-key. Forward it as an X-API-Key header; the Worker passes it straight through to api.endoflife.ai.

Notes

• Same-zone trap: api.endoflife.ai is a Worker on this zone; a plain fetch() to it goes to origin and fails. The API service binding is mandatory in prod.
• Static feeds are fetched from the site origin with cf.cacheTtl = 600 and memoised in the isolate for 10 minutes.
• Per-client rate limits: relies on the upstream API's limits. If MCP traffic grows, add a KV rate-limiter keyed on CF-Connecting-IP before the upstream call.
• Discovery card is also served from the main site at https://endoflife.ai/.well-known/mcp/server-card.json.