talkops-ai/talkops-prometheus
Operational skills and MCP execution server for Prometheus 3.0+, Thanos, Mimir, Cortex, and VictoriaMetrics: PromQL authoring, RED/SLO triage, trace exemplar pivoting, rule simulation & testing, TSDB cardinality FinOps, Kubernetes Operator scrape CRDs, and exporter lifecycle management.
Authors, validates, unit-tests, backtest-simulates, and non-destructively deploys or deletes Prometheus recording and alerting rule groups across Kubernetes PrometheusRule CRDs, HTTP Ruler APIs, local rule files, and GitOps YAML. Always read SKILL.md first before calling run_command, and inspect CRDs via Prometheus MCP resources rather than raw kubectl. Use when creating, auditing, or tuning Prometheus alerting/recording rules, fixing flapping or chronic alerts, diagnosing unpicked PrometheusRule CRDs, running promtool unit tests, simulating historical/synthetic rule firing, or updating/deleting rule groups. Do not use for ad-hoc PromQL exploration without rules (use authoring-promql-and-analyzing-metrics), Grafana Loki LogQL Ruler rules (use authoring-loki-alerting-and-recording-rules), Alertmanager routing/silences (use validating-alertmanager-routing or managing-alert-silences), or Argo Rollouts AnalysisTemplates (use gating-rollouts-with-analysis-and-experiments).
Discovers Prometheus metric schemas, authors and lints type-safe PromQL 3.0 queries (classic and native histograms, UTF-8 OTel selectors, vector joins, subqueries, and zero-fill ratios), and analyzes instant, range, or period-over-period time series. Use when asked to write, lint, fix, or execute a PromQL query, explore available metrics and labels, calculate percentiles or rates, compare current vs historical baseline windows, or render time-series charts. Do not use for authoring or deploying PrometheusRule CRDs and alert unit tests (use authoring-and-testing-prometheus-rules), incident RED/SLO burn-rate triage (use triaging-prometheus-incidents-and-slos), TSDB cardinality FinOps relabeling (use optimizing-tsdb-cardinality-and-finops), LogQL log queries (use authoring-logql-and-analyzing-logs), or TraceQL queries (use authoring-traceql-and-metrics-queries).
Configures, validates, troubleshoots, and manages Prometheus scrape discovery across Kubernetes Operator CRDs (ServiceMonitor, PodMonitor, Probe, ScrapeConfig), Prometheus 3.0 native histogram and OTLP ingestion, and File SD target groups. Always read SKILL.md first before calling run_command, and inspect CRDs via Prometheus MCP resources rather than raw kubectl. Use when onboarding an existing Kubernetes Service or Pod metrics endpoint for scraping, configuring Blackbox Probe or external ScrapeConfig CRDs, managing file_sd JSON targets, or debugging missing/DOWN scrape targets. Do not use for installing third-party exporter workloads like postgres_exporter or redis_exporter (use onboarding-exporters-and-instrumentation), TSDB cardinality drop relabeling (use optimizing-tsdb-cardinality-and-finops), or provisioning OpenTelemetry Collectors (use provisioning-opentelemetry-collectors).
Selects, renders, deploys, verifies, and uninstalls Prometheus infrastructure exporters across the 20-exporter catalog (postgres_exporter, redis_exporter, kafka_exporter, blackbox_exporter, node_exporter, kube-state-metrics, nginx_exporter, jmx_exporter, mysqld_exporter, aws_cloudwatch_exporter, etc.) and guides Python/Go/Java/Node.js/Rust/.NET Prometheus SDK instrumentation with OpenMetrics trace exemplars. Use when deploying or removing a Prometheus exporter for databases/caches/brokers/nodes, rendering offline GitOps exporter manifests with DSN secret redaction, fixing Python multiprocess counters, or wiring OpenMetrics trace exemplars. Do not use when a workload already exposes /metrics and only needs a ServiceMonitor (use configuring-prometheus-scraping-and-discovery), for OTel Operator zero-code injection (use instrumenting-kubernetes-workloads), for Helm database chart installs (use deploying-helm-charts), or for OTel Collector provisioning (use provisioning-opentelemetry-collectors).
Audits Prometheus TSDB series cardinality, label value fan-out, CronJob pod churn, histogram bucket bloat, and unreferenced high-cost metrics; sizes TSDB RAM/SSD/sharding capacity; and generates safe metricRelabelings, recording rules, hashmod sharding, and remote_write FinOps configurations. Use when reducing Prometheus memory or Mimir/Thanos cloud bills, investigating high series counts or label fanout, pruning histogram buckets, or sizing/sharding Prometheus. Do not use for general PromQL query writing (use authoring-promql-and-analyzing-metrics), Grafana Loki log stream cardinality (use optimizing-loki-cardinality-and-finops), OTel Collector OTTL attribute filtering (use controlling-telemetry-cardinality-and-noise), or basic ServiceMonitor onboarding (use configuring-prometheus-scraping-and-discovery).
Triages Kubernetes service incidents, latency regressions, USE saturation, firing Prometheus alerts, and Google SRE multi-window SLO error-budget burn rates using RED golden signals, period comparisons, and OpenMetrics trace exemplars. Use when investigating an outage or page, evaluating 30-day SLO error-budget burn and hours to exhaustion, generating multi-window SLO burn-rate rules, diagnosing RPS drops vs target-down blind spots, or pivoting from latency/error spikes via exemplars. Do not use for clustering Alertmanager alert storms or creating Alertmanager silences (use triaging-alertmanager-alerts or managing-alert-silences), LogQL log summarization (use triaging-loki-incidents-and-errors), or TraceQL span-cohort diffs (use triaging-tempo-service-errors).