Pardinus AI Suite built in, for Checkmk

The intelligence layer for Checkmk.

Your monitoring already knows something is wrong. Pardinus works out why, with the evidence it used, a confidence score, and a next step that waits for your approval. Inside Checkmk, on the model you choose.

Delivery

A Checkmk MKP. No extension, no second console.

Model

Bring your own. Any OpenAI compatible endpoint.

Actions

Nothing runs without a person pressing Run.

Cost

Token spend tracked per site, user and operation.

The problem

Six alerts. One cause.

A saturated host, a slow database, a failed checkout, and a memory ceiling on a node. Monitoring reports them as independent problems, at 3am, to one person.

Pardinus reads them together, with the history of everything that looked like this before, and returns one investigated incident.

Host down srv-app-04 CPU high 94% / 5m avg Memory critical node-k8s-11 Application failed checkout-api Database slow pg-primary Root cause confidence 0.87

What's in the suite

Nine modules, one add-on

Pardinus started as chat and root cause analysis. It is now the operational layer around them: what to check first, what to run, what to learn, what it costs, and who approves.

Investigate

A verdict with confidence, the Checkmk facts behind it, and the specialist agent that produced it.

Quick Diagnose

A read only playbook per alert type. One known command, raw output, no tokens spent.

Action Path

Suggested remediation you edit, select, and approve. Every run over SSH, every run audited.

Proactive RCA

Analysis that starts the moment a service goes critical, scoped by a normal Checkmk ruleset.

Trend Analysis

Recurring patterns, services that occur together, and tuning candidates, computed before anything is written.

Knowledge Base

Findings become reusable knowledge only when a human approves them, with a note on why it happens.

Configuration Auditor

A risk score for the site: shadowed rules, duplicates, missing thresholds, and general hygiene.

Chat

Ask about the estate from any dashboard, in your language, with per user history.

Health & Finance

Self monitoring and token cost per site, per operation, per user, with month end projection.

Investigate

Every problem, routed to the right specialist

The live queue shows host and service problems with the state of their analysis: already analysed, picked up automatically, or still untouched. Routing is automatic across Linux, Windows, database and network, and you can override it per alert.

The queue tells you where you stand

Recurring alerts are marked as recurring, with the count. An alert that already has an answer offers the answer instead of another analysis.

  • Auto RCA, on demand RCA, or none, visible per row
  • Repeated unchanged alerts reuse the previous result
  • Filter by site, host, folder or service pattern
Pardinus live alert queue in Checkmk showing service problems with RCA status per row
Live alert queue
Pardinus RCA report showing verdict, medium confidence 0.60, evidence count and agent routing notes
RCA report - verdict, confidence, evidence

A verdict you can argue with

The report names the likely cause, scores its own confidence, and lists the Checkmk facts it used. Where the evidence is thin, it says so rather than filling the gap.

  • Confidence score with alternatives and data gaps
  • The evidence package behind the answer, in full
  • The performance graph for the service, in the report
  • Run again against a different specialist agent

In this run it found no incident at all: the critical threshold was set at 2%, actual usage was 4.39%, and the alert had flapped 21 times in a day.

Answer first, spend later

Not every alert deserves a model

Most on call questions are answered by one command. Quick Diagnose runs it as a read only, deterministic check in about a second and offers to escalate only if what you see isn't enough.

Quick Diagnose

A named playbook attached to the alert type: top processes by CPU, top memory consumers, filesystem growth. Raw output, exit code, timing.

  • Read only by definition, nothing changes on the host
  • No tokens consumed
  • One button to escalate into a full analysis
Pardinus Quick Diagnose showing top 10 memory consumers as read only command output
Quick Diagnose - read only
Pardinus Action Path with suggested remediation commands awaiting approval, above the Checkmk performance graph
Action Path - remediation with approval

Action Path: the Run button is the approval

When an analysis suggests a fix, the steps arrive as an editable list. You select, reorder, edit, and run. Commands execute over SSH with no shell, no pipes, no globs, and anything outside the allowlist has to be requested and granted first.

  • Per host concurrency limits and loop detection
  • Credentials from the Checkmk password store
  • Every execution written to the audit log

Native

Configured in Checkmk, not beside it

Scoping happens where your team already scopes everything else. Proactive analysis and the SSH accounts it uses are ordinary rulesets, with folders, host tags, labels and explicit hosts, and the usual first match wins behaviour.

Rulesets, not a parallel config

Choose which services start an analysis on their own, on which state change, with which cooldown. Choose which machines use which login. Nothing to learn that isn't already Checkmk.

  • Proactive analysis scoped by service rule
  • Service accounts scoped by host rule
  • Shadowed rules detected and explained, not silently ignored
Checkmk ruleset editor for Pardinus Proactive RCA on services
Setup - Proactive RCA ruleset
Pardinus service account settings listing the Checkmk rules that select it, including a rule that never fires
Service accounts - shadowed rule detected

Learn

The second time should be cheaper than the first

Statistics run on your incident history on a cadence: high volume alerts worth tuning, services that always fail together, patterns that keep coming back. The numbers are computed deterministically; the model only writes the explanation.

Nothing enters the knowledge base on its own

Proposals wait in a queue with their support count and confidence. You approve or reject, optionally adding the reason it happens. Both answers train the next analysis.

  • Tuning candidates for alerts that fire too often
  • Services that occur together across sites
  • Recurring canonical patterns with support and confidence
Pardinus Trend Analysis with knowledge base promotions awaiting approval
Trend Analysis - promotions awaiting approval
Pardinus Configuration Auditor showing a site risk score and categorised findings
Configuration Auditor

And the configuration itself gets reviewed

The Configuration Auditor scores the site and lists what is wrong with it: duplicate rules, shadowed rules, rules that never match, thresholds that are missing or implausible. Export as Markdown or JSON.

  • Risk score with critical, warning and informational findings
  • Site performance advisor and cleanup suggestions
  • Audit history, so you can show the trend

Ask

Talk to your estate, from any dashboard

Context-aware answers

The assistant sits beside the view you are already looking at. Ask about a host, a service, a dashboard you don't recognise, or a Checkmk concept, and get an answer about your systems, not a documentation quote.

  • Side panel on any dashboard, or a full page with history
  • Answers in the language you ask in
  • Offers to open a full analysis when it finds something
  • Voice input and spoken replies
Pardinus assistant panel open beside a Checkmk Windows hosts dashboard
Assistant - in dashboard

Control

Answerable to somebody

An AI near production monitoring needs a shorter leash than a dashboard does. Permissions, spend and behaviour are configuration, not promises, and when an agent wants to do something it isn't allowed to do, it asks and waits.

Governance is a tab, not a footnote

  • Capability grants queued for human approval
  • Execution policy and command allowlists
  • Cost limits per site and per period
  • Per role permissions inside Checkmk
  • Redaction of sensitive values before anything leaves
Pardinus settings on the Governance tab showing capability grants awaiting approval
Settings - Governance
Pardinus Finance dashboard showing LLM spend by site and projected month end
Finance - spend by site
Pardinus Health page showing provider status, latency and per site reachability
Health - distributed monitoring

What it costs, and whether it is healthy

Token spend is tracked per site, per operation and per user, with a month end projection and a plain answer to the awkward question: how much of this spend bought nothing, because the alert had already been analysed.

  • Spend month to date against projection, by site
  • Tokens saved by reusing results for repeated alerts
  • Provider reachability, success rate and latency percentiles
  • For distributed setups: whether every site can reach a provider

Deterministic before AI

Context is built from Livestatus and the REST API first. The model interprets a package it did not assemble.

Deterministic fallback

If generation fails, the evidence package is still returned, so the shift keeps moving.

Bring your own model

Any OpenAI compatible endpoint: self hosted, corporate, Azure OpenAI, or a public provider. Your data boundary.

Runs where the host is monitored

In distributed setups the analysis executes on the site that owns the host, not through a central bottleneck.

API first

Wire it into what you already run

Built for incident pipelines

Every capability is exposed through a versioned REST API. Trigger an analysis from your incident pipeline, pull the report into the ticket, attach the evidence to the channel where the shift actually talks.

n8n PagerDuty ServiceNow Jira Slack MS Teams Webhooks
POST
/api/v1/rca/run

Start an analysis

GET
/api/v1/rca/{job_id}

Poll status

GET
/api/v1/rca/{job_id}/report.json

Fetch the report

POST
/api/v1/chat/query

Ask a question

Bearer auth. Callback or polling. Roughly fifteen minutes to attach an analysis to every new incident.

Pilot

See it against your own alerts

A guided pilot on your Checkmk, your model endpoint, your noisiest services. If it doesn't reduce the time your team spends deciding what to look at first, it isn't worth installing.