mcp-uptime

Is that MCP endpoint alive, what does it actually expose right now, and has its tool schema changed since you last looked?

52% of remote MCP endpoints are dead. Schema drift is worse than downtime, because downtime is loud and a quietly-changed inputSchema fails mid-agent-run and looks like a model error. This service answers four questions about any public MCP endpoint: liveness, tool inventory, schema drift, and 30-day uptime.

It ships twice from one codebase:

Surface	What it is	Why
Cloudflare Worker	An MCP server (/mcp) plus a plain HTTP API (/v1/*)	Reach — agents discover it through the MCP registry
Apify Actor	Same check logic, pay-per-event	Metering + payout — Apify counts runs and users natively

Free tier: 20 checks/day, on-demand. Paid tier ($9/mo, not wired up yet): scheduled monitoring, drift history, webhook alerts.

---

MCP tools

Tool	Probes?	Returns
check_mcp_endpoint	yes	status, transport, latency, tool count (or the full list), schema hash, drift
check_schema_drift	yes	tools added / removed / modified since the last snapshot
get_uptime	no	uptime % and incident windows over the last N days (max 30)
get_known_tools	no	last observed tool list + when the schema last changed

Point any MCP client at POST https://<your-worker>.workers.dev/mcp. Streamable HTTP; no auth needed for the free tier.

// tools/call → check_mcp_endpoint
{
  "endpoint": "https://mcp.example.com/mcp",
  "status": "up",              // up | degraded | down
  "transport": "streamable-http",
  "latencyMs": 214,
  "protocolVersion": "2025-06-18",
  "server": { "name": "example-mcp", "version": "1.4.0" },
  "toolCount": 12,
  "toolsHash": "9f2c…",
  "drift": { "changed": true, "added": [], "removed": ["search_v1"],
             "modified": [{ "name": "search", "fields": ["inputSchema"] }] },
  "firstSeen": false,
  "error": null
}

degraded is its own status on purpose: the server answered initialize but its tools/list failed. Collapsing that into down would hide exactly the failure this product exists to surface.

HTTP API

Method	Path	Auth	Notes
POST	/v1/check	none	{ "url": "...", "include_tools": false }
GET	/v1/status?url=…&days=30	none	stored history, no probe
GET/POST	/v1/monitors	key	list / create (scheduled checks + webhook)
GET/DELETE	/v1/monitors/:id	key	own monitors only
POST	/mcp	none	MCP Streamable HTTP
GET	/health	none	liveness
GET	/admin/stats	operator secret	probe call counts; 404s without the secret

Security model

The whole job of this service is to fetch a URL a stranger hands it, so the guard rails are the product, not decoration. Implemented against the access-control matrix in ../PROBE-SPEC.md §SECURITY:

• SSRF guard (src/core/ssrf.ts) — http/https only; loopback, RFC1918, CGNAT, link-local (incl. 169.254.169.254), NAT64, multicast, reserved, and IPv4-mapped IPv6 all rejected; internal TLDs and bare hostnames rejected; embedded credentials rejected; internal service ports blocked. Re-applied to every redirect hop — redirects are followed manually because a public host is allowed to bounce you at metadata.
• Rate limit before auth — a coarse per-IP limiter runs before any key lookup, so an unauthenticated flood never reaches the tenants table. Plan quotas apply afterwards, on the resolved principal.
• Fail closed — if the counter store errors, requests are denied (503), not admitted.
• No cross-tenant reads — monitor queries are tenant-scoped in SQL. Another tenant's monitor returns 404, not 403, so ids are not an existence oracle.
• No secrets or caller data in logs — callers are stored as SHA-256 hashes, never raw IPs; API keys are stored only as hashes; error bodies are fixed codes, never remote content or internals; URL query strings (which routinely carry tokens) are dropped before an endpoint is persisted.
• Untrusted remote content — tool names and descriptions come from the endpoint being probed and are handed to agents, so they are count-capped (200) and length-capped before they reach storage.
• Dependencies — exact-pinned, with a lockfile. The Worker has zero runtime dependencies: the MCP server is hand-rolled JSON-RPC.

Verified by tests/ssrf.test.ts, tests/ratelimit.test.ts, and tests/tenant.test.ts — the three tests the spec names.

Layout

src/core/     runtime-agnostic: SSRF guard, prober, drift, uptime, quotas, auth
src/worker/   Cloudflare Worker — HTTP router, MCP server, D1 adapter
src/actor/    Apify Actor entry point
tests/        node:test, no test framework to install
.actor/       actor.json, input schema, Dockerfile, PPE event reference
schema.sql    D1 schema

src/core imports nothing but the Web platform (fetch, URL, crypto.subtle), which is why both wrappers run identical logic and the tests exercise the real thing rather than a stand-in.

Development

npm install
npm test          # 37 tests, no network, no framework
npm run typecheck
npm run dev       # wrangler dev on http://localhost:8787

Tests run the TypeScript sources directly under Node's strip-only mode (Node 22+), so there is no build step and no test runner to install. That constrains the source: no enums, no namespaces, no constructor parameter properties.

Deployment: see DEPLOY.md. Accounts and go-live: see LAUNCH-README.md.