Grafana MCP: dashboards, metrics and logs for agents

When an alert fires, the on-call engineer opens Grafana to find the dashboard, writes PromQL against Prometheus, digs through Loki, and ten minutes go by. Grafana MCP lets an agent run those queries through your Grafana instance: search dashboards, query metrics and logs, check alerts and on-call. Grafana handles data source authentication, so credentials never pass through the MCP server. Grafana Labs maintains it, and it had about 3.5k stars as of 2026-10-06.

What it does

• Dashboards: search by title, folder or tag. get_dashboard_summary returns a compact overview and get_dashboard_property pulls only the parts you name via JSONPath, so a full dashboard JSON doesn’t swamp the context.
• Metrics and logs: run PromQL (including histogram percentiles) and LogQL, read labels and metadata, and see log patterns Loki has detected.
• More data sources: opt-in tools for SQL sources such as ClickHouse, Snowflake and PostgreSQL, plus CloudWatch, Elasticsearch, InfluxDB and Graphite.
• Alerting and on-call: alert rule status and notification policies, Grafana OnCall schedules and who’s on call now, and search, create or update Grafana Incident incidents.
• Category switches: --disable-write drops every write tool, and --disable-<category> removes a whole category.

Who it’s for

• SREs who want the agent to gather the relevant metrics, logs and alerts before they start an investigation.
• Developers on Grafana who want the agent to check production metrics after shipping a change.

Setup

Needs Grafana 9.0 or later, a service account token and uv:

{
  "mcpServers": {
    "grafana": {
      "command": "uvx",
      "args": ["mcp-grafana"],
      "env": {
        "GRAFANA_URL": "http://localhost:3000",
        "GRAFANA_SERVICE_ACCOUNT_TOKEN": "<your service account token>"
      }
    }
  }
}

Grafana Cloud users swap in their instance URL. The README also covers Docker, binaries and Helm.

Our take

Grafana holds a special spot among observability servers: many teams already route Prometheus, Loki and Elasticsearch through it, so one server reaches most of their monitoring data without a separate MCP per backend. It’s careful with context, too, with dashboard summaries, JSONPath lookups and truncated results all aimed at that. Note that write tools are on by default: it can edit dashboards, create and delete alert rules, and open incidents. Add --disable-write for query-only use and give the service account minimal permissions. Anonymous usage statistics are on by default and report which tools were called to Grafana Labs every four hours; --usage-stats=disabled or DO_NOT_TRACK=1 turns that off. Licensed Apache-2.0.