The short answer: if your home AI server's endpoint answers an unauthenticated request from the internet and speaks MCP, it is exposed — anyone who finds the URL can talk to it. HomeShield's free check answers the question with a real probe, not a guess: paste the endpoint's public URL and it grades the response EXPOSED (HTTP 200 with MCP markers), REACHABLE (200, no markers), COVERED (an authentication response: 401, 403, or 407), or UNKNOWN — and separately grades the tool surface the server shows the world. Probe behaviour as shipped in HomeShield's prober, November 2026.
It makes safe, read-only requests and reads what comes back — the same thing anyone on the internet can do. One scan checks your declared URL plus the common MCP paths on the same host (/mcp, /sse, /api/mcp, /messages), the slash and http variants, the certificate, and the cross-origin settings. Then the classification is mechanical:
| What answers | Verdict | What it means |
|---|---|---|
| 200, and the response speaks MCP (protocol markers or an SSE stream) | EXPOSED | Your server is talking to strangers. Authenticate it or take it off the public internet. |
| 401, 403, or 407 | COVERED | Authentication is doing its job — the correct posture for a reachable home AI server. |
| 200, but nothing MCP-shaped | REACHABLE | Something is there — a web UI, another app, or MCP on an unguessed path. A look-closer, not a pass. |
| 404, a server error, or an uninterpretable answer | UNKNOWN | No honest signal. The report says which, and why. |
Almost never by malice and rarely by a hack. The three usual routes: a port forward added to the router so the owner could reach the server from outside, then forgotten; a tunnel (Cloudflare Tunnel, Tailscale Funnel, an ngrok-style link) created for convenience — HomeShield's probe detects tunnel setups from the outside, because a tunnel removes the port forward, not the exposure; and a server that binds to every network interface by default on a machine that is itself reachable. If you are not sure which route yours took, you are the normal case — that is what the check is for. New to the vocabulary? What an MCP server is explains the thing being exposed.
EXPOSED: treat it as this week's problem. Put authentication in front of the server or close the public path today; then shrink the tool surface (the report grades it None seen / Low / Medium / High) and re-run the check to watch the verdict change. COVERED: keep the authentication, and work the home-lab checklist — covered is a posture, not an audit. REACHABLE: identify what actually answered; if it is your server's web UI, it deserves the same authentication question. UNKNOWN: verify the exact URL and re-run; the report names the reason it could not classify.
And the honest boundary, stated on every report: this reads one declared public endpoint from outside. It cannot see your LAN or your router — the local-scan guide covers the inside of the house.
Run HomeShield's free check on its public URL. The probe reads the actual response: a 200 with MCP markers grades EXPOSED, a 200 without markers grades REACHABLE, an authentication response (401, 403, 407) grades COVERED, and anything else grades UNKNOWN with the reason stated.
The endpoint answered with HTTP 200 and spoke MCP (protocol markers or an SSE stream) to an unauthenticated internet request. Anyone who finds the URL can talk to your server. Put authentication in front of it or take it off the public internet.
An authentication response (401, 403, or 407) grades COVERED — the correct posture for a home AI server that must be reachable. Covered is not the same as audited: the check reads the posture from outside and cannot see your LAN, which the local-scan guide covers.
No. The probes are safe and read-only: the declared URL plus the common MCP paths on the same host (/mcp, /sse, /api/mcp, /messages), the slash and http variants, the certificate, and the cross-origin settings. Nothing is written, exploited, or stored beyond the report.
Anonymous scans run 3 per day; a free account gets 10 per month. The check itself is the free core job — no signup is needed to run it.
HomeShield MCP uses cookies to run the site and, through Google AdSense, to serve ads. Accept or decline non-essential cookies. Details are in the privacy policy.