Server
zabbix-mcp — vendored third-party, read-only, three tools. See
zabbix-metrics-history for the shared cautions.
Current problems vs history — use the right one
| Question | Method | Note |
|---|---|---|
| What is wrong now? | problem.get |
Reads a dedicated problem table. This is the right one |
| What happened then? | event.get |
Historical, heavier, needs a time window |
| What rule fired it? | trigger.get |
The definition behind a problem |
zabbix_api("problem.get", {
"output": "extend", "selectAcknowledges": "extend", "selectTags": "extend",
"recent": false, "sortfield": ["eventid"], "sortorder": "DESC"
})
Every problem you report carries: severity · host · when it started · how long it has been active · acknowledgement state. "How long" is the question NetClaw could not answer at all before this integration existed — do not drop it.
The distinction that matters most here
"No active problems" and "the NMS could not be reached" are not the same answer.
An empty problem list is a legitimate, positive finding — the monitoring system looked and found nothing wrong. An unreachable NMS is a failure to look.
They are both an empty result over the wire, and reporting the second as the first is the most misleading thing this skill can do: it tells an engineer everything is fine at exactly the moment monitoring is blind.
Three outcomes, three different sentences:
| Say | |
|---|---|
| NMS reachable, nothing wrong | "Zabbix reports no active problems as of <time>." |
| NMS unreachable | "Zabbix could not be reached — this is not a statement about the network." |
| Credentials rejected | "Zabbix rejected the credentials — the monitoring state is unknown." |
Acknowledgement is a workflow fact, never a resolution
An acknowledged problem is still happening. Someone has said "I've seen this", which is a fact about the team, not about the network.
Report it as "active, acknowledged by <who> at <when>". Never let acknowledgement soften the description of the underlying condition, and never let a filtered-out acknowledged problem disappear from a count.
Filtering
Filter before you answer — by severity, host or group — rather than returning everything and asking the
reader to ignore rows. Use severities, hostids or groupids on the call.
Severities: 0 not classified · 1 information · 2 warning · 3 average · 4 high · 5 disaster.
When you filter, say what you filtered out, so "two problems" is never mistaken for "two problems exist".
Resolved problems
event.get with value: 0 gives resolutions. Report both onset and resolution times — a problem that
lasted four minutes and one that lasted four hours are different incidents, and the duration is usually the
point of the question.
Boundaries
| Want to… | Use |
|---|---|
| Metric values over time | zabbix-metrics-history |
| Availability and inventory | zabbix-availability |
| Unsolicited traps | snmptrap-mcp — push, not poll |
| Flows | ipfix-mcp |
| Instrumented metrics / dashboards | prometheus, grafana |
| SaaS monitoring | auvik, thousandeyes, datadog |
| Current device state | pyats, multivendor-cli, fortinet — this is the poller's view over time |
| Acknowledge, close or suppress anything | nothing here. This integration is strictly read-only |
Rules
- Never report an unreachable NMS as "no problems".
- Always give duration, not just onset.
- Acknowledged ≠ resolved. Say "acknowledged", never imply "handled".
- Say what you filtered out.
- Read-only. Acknowledging an alert is a real action a human owns; there is no write path here.