Skip to main content
Ranking6 min read

Top 10 MCP servers for DevOps in 2026

From Kubernetes triage to Terraform plan review: the ten MCP servers that collapse the time between “incident fires” and “incident resolved”.

DevOps is the killer domain for MCP — the work is mostly reading heterogeneous systems and making small, careful writes. Here are the ten MCP servers that should be in every SRE’s loadout in 2026.

1. Kubernetes MCP

kubectl-grade access for agents. Inspect pods, stream logs, roll back deployments. RBAC-scoped to whatever the kubeconfig allows — use a read-only service account for day-to-day.

2. Terraform MCP

Read state, preview plans, inspect module outputs. “Why is this diff showing — did someone modify resources out-of-band?” answered in one turn.

3. AWS MCP

Cross-service triage: S3 inventory, Lambda invocations, CloudWatch log queries, RDS status. Use a read-only IAM role by default.

4. Grafana MCP

Dashboards, alerts, panel data. Ask “show me the p99 latency for the auth service in the last hour”; get a number without logging into the UI.

5. Datadog MCP

Monitors, logs, APM, incidents. Especially useful during on-call — the agent triages logs while you focus on the fix.

6. PagerDuty MCP

Incident status, on-call schedules, escalation policies. “Who is on-call for the auth service right now?” –– instant answer.

7. Argo CD MCP

GitOps deployments status. Inspect diffs between cluster and git, trigger sync, roll back — safely.

8. Cloudflare MCP

Workers, KV, R2, DNS, WAF rules. For incident response: flip a rate-limit rule without logging into the dashboard.

9. HashiCorp Vault MCP

Retrieve secrets on demand, list policies, audit leases. Especially important when an agent needs temporary credentials for a one-off task.

10. Sentry MCP

Error triage during rollouts. “Any new error signatures since the last deploy?” — agent diffs the releases and tells you.

Example: the 3 AM incident loop

With this loadout, a 3 AM page collapses to:

  1. Ask agent: “what changed in the last hour?” → Sentry + Datadog tell you error rate tripled for service X.
  2. “what deployed to X recently?” → Argo CD shows the git SHA; GitHub MCP shows the PR.
  3. “roll back” → Argo CD reverts.
  4. “confirm error rate dropped” → Datadog confirms.

Four chat turns, no dashboard hopping.

Safety notes

DevOps MCP servers have the widest blast radius. Defaults matter:

  • Kubernetes — read-only by default; a write-capable instance only in a locked-down host config.
  • AWS — scoped IAM, no iam:*, no production write access.
  • Cloudflare / DNS — explicit approval required for every write call.
  • Vault — short-lived tokens only; audit log every access.

Where to go next

Loadout

Build your AI agent loadout

Directory
Contact
© 2026 Loadout. Built on Angular 21 SSR.