Grafana

Grafana's official MCP bridges your observability stack into Claude or Cursor. It reads dashboards, datasources, and active alerts, and lets you run PromQL and LogQL queries without leaving the editor. The pitch is simple: instead of alt-tabbing to Grafana during an incident, you ask your agent to summarise the last hour of alerts or pull recent error logs from Loki. It works. I've used it most during postmortems. You can ask Claude to cross-reference a Sentry error ID with Loki logs, then draft a timeline based on dashboard snapshots and alert history. The agent has schema awareness for Prometheus metrics, so it can suggest PromQL queries that actually match your label structure. That's more useful than it sounds when you're three hours into debugging and can't remember if your service uses `job` or `service_name`. Setup is straightforward if you already have a Grafana instance and an API key with viewer permissions. The MCP exposes read-only tools, so it won't accidentally silence alerts or delete dashboards. You point it at your Grafana URL, drop the key into your environment, and it surfaces everything through natural language. The main quirk is that it's read-only. You can't create dashboards, modify queries, or acknowledge alerts. That's by design, but it means you're still opening the web UI for anything beyond inspection. The other limitation is host support. It works cleanly in Claude Desktop, but Cursor and VSCode need manual config tweaking. If you're on a different editor, you're writing your own wrapper. This is for teams already running Grafana in production. If you're not using Loki or Prometheus, half the tooling is wasted. If you are, and you spend time in Claude Desktop drafting runbooks or analysing incidents, this is worth the five-minute install. It won't replace your observability workflow, but it tightens the loop between asking a question and getting an answer.