MCP server monitoring
Uptime monitoring asks whether a server responds. MCP server monitoring asks a harder question: does it still honor the contract your agents were built against? A server can be perfectly up while silently renaming a field, narrowing an enum, or rewriting the description that steers your model's behavior.
Zevruna polls each monitored server on your schedule (daily on the free tier down to every minute), performs the standard MCP handshake, canonicalizes and hashes the advertised contract, and diffs every change against the version your manifests pin. Unchanged polls store only a heartbeat; the record stays lean.
Alerts that name the blast radius
A breaking change fires within one poll cycle, and the alert doesn't just say a server changed, it says which of your agents are affected, which exact fields, and in which files and lines, with a suggested patch. Agents that don't touch the changed tool stay silent. Signal, not noise.
Detection without traffic
Error trackers need production traffic to hit the broken path; low-traffic tools break on Friday and surface Monday. The poller catches drift whether or not anyone called the tool, median detection on the public monitor is measured in minutes, not days.
What one poll actually does
The monitor performs the standard MCP handshake and asks three questions: tools/list, resources/list, prompts/list. It never invokes a tool with real arguments, so monitoring cannot charge your metered endpoints or write to anything.
What comes back is canonicalized, keys sorted, volatile fields stripped, so that two identical contracts hash identically regardless of how either server happened to serialize them. That hash is compared to the last one. Around ninety-five percent of polls match, and those store a heartbeat rather than a snapshot, which is what keeps a decade of history cheap instead of enormous.
From alert to a fix you can prove
The alert names the affected agents and fields. Accepting the new contract repins the affected manifests, which makes CI fail until the call sites match, and the green check is what closes the incident. Nobody marks it resolved by hand, and the exposure window between detection and that check is recorded as evidence.
Questions
Does monitoring call my tools?
Never with real arguments. Cloud monitoring is handshake-only: tools/list, resources/list, prompts/list. Live-invocation assertions exist but are opt-in and customer-authored.
How fast is detection?
One poll cycle after the change lands: daily on Free, fifteen minutes on Pro, one minute on Scale.
What if the server is down rather than changed?
Those are recorded as different things. A failed handshake is a reachability failure and does not produce a contract diff, because an unreachable server has not told us its contract changed, only that we could not ask.
Do you need credentials for a private server?
Only if the server requires them to complete a handshake. Auth headers are stored encrypted and used for nothing but the three list calls above.
How is this different from uptime monitoring?
Uptime asks whether the server responds. This asks whether it still honours the contract your agents were built against. A server can be green on every uptime check while silently renaming the field your agent sends.
Paste any endpoint on the homepage for an instant, free contract report.