MCP Server Health Monitoring

What it checks

When you monitor an MCP (Model Context Protocol) server, Merlonix does not just ping a URL — it runs a real MCP handshake on your check cadence: a JSON-RPC initialize followed by tools/list, exactly the way an AI agent (Claude, Cursor, or your customers' agents) connects. It records whether the handshake succeeded, the negotiated protocol version, the handshake latency, and the set of tools the server exposes.

This matters because a generic uptime check can miss an MCP outage entirely: the endpoint URL may still return HTTP 200 while the MCP handshake fails, so every agent that tries to use it breaks even though a normal monitor stays green.

When it alerts

• Down (critical): the server stops responding to the handshake — a failed initialize, a transport error, or a non-responding endpoint. You get one open alert per asset while it is down; it re-fires only after a recover→down cycle, so you are not paged every tick.
• Recovered (info): the server answers the handshake again after being down.
• Tool / capability / protocol drift (the silent breaking change): when both the previous and current checks handshaked successfully but the tool contract changed — a tool removed or renamed, a dropped capability, or a protocol-version change — Merlonix flags it. A removed or changed tool breaks agents mid-flight with no outage a URL monitor could see.

If the endpoint is unreachable or the handshake is indeterminate, the check is recorded and never falsely alerts as drift.

Turning it on

Add an asset with an MCP endpoint URL and enable MCP server health monitoring. To check any public MCP endpoint right now with no signup, use the free MCP server health checker on the marketing site. See the Billing page for which plan covers the assets you want to watch.