MCP Grafana

What grafana/mcp-grafana MCP server does

grafana/mcp-grafana MCP server gives MCP clients access to Grafana dashboards, datasources, alerting, incidents, annotations, OnCall information, and selected observability analysis workflows. It can retrieve a complete dashboard by UID, or return a smaller summary or selected properties when full dashboard JSON would use too much context. Dashboard tools also expose panel queries and datasource details, and the server supports dashboard creation, updates, patches, annotations, folders, snapshots, deeplinks, and rendered panel images.

Operational tools cover alert rules and alert routing separately. Agents can inspect or change rule configuration, compare rule versions, and filter rules by state, folder, group, or name. Routing tools cover notification policies, contact points, and time intervals. Incident tools can create or retrieve incidents and add timeline notes, while OnCall tools provide alert group, schedule, shift, and currently on-call user information.

How it works

The server runs locally as an MCP process and communicates with Grafana using the configured instance URL and service account token. Grafana remains the source of datasource credentials for supported backends such as Snowflake; the README states that those credentials are not exposed to the MCP server.

Query and investigation capabilities span Prometheus, Loki, InfluxDB, ClickHouse, CloudWatch, Graphite, Athena, Snowflake, Elasticsearch, and OpenSearch, subject to the relevant Grafana datasource and enabled tool group. It can query metrics, logs, traces, and datasource metadata, and includes targeted analysis for Loki error patterns and Tempo slow requests. It can also generate Explore links with embedded time ranges and queries.

Setup and configuration

The recommended configuration uses uvx with the mcp-grafana package. Set GRAFANA_URL to the Grafana address and GRAFANA_SERVICE_ACCOUNT_TOKEN to a service account token. Grafana Cloud instances use their instance URL in place of a local address. The README also documents Docker, binary, and Helm installation paths, but the MCP client example uses uvx.

The server requires Grafana 9.0 or newer for full functionality. Add runpanelquery to the --enabled-tools setting to enable panel query execution. Query examples require examples; InfluxDB, ClickHouse, CloudWatch, Graphite, Athena, Snowflake, and Elasticsearch/OpenSearch each have their own enablement value.

Tools and capabilities

Key capabilities of grafana/mcp-grafana MCP server include:

• Summarizing dashboards or selecting dashboard fields with JSONPath.
• Reading panel queries, datasource metadata, annotations, and dashboard images.
• Creating folders, annotations, snapshots, incidents, and datasources.
• Managing alert rules and notification routing.
• Checking datasource health and querying supported datasource types.
• Inspecting Loki labels, log patterns, Prometheus metadata, CloudWatch resources, and database schemas.
• Finding slow requests in Tempo and elevated error patterns in Loki.
• Retrieving Grafana Sift investigations and analyses.

Datasource query tools support datasource-specific features such as PromQL, LogQL, InfluxQL or Flux, SQL, Graphite render queries, Lucene or Elasticsearch Query DSL, and Grafana macro or variable substitution where documented.

Limitations and notes

Large dashboards can consume substantial context because complete dashboard retrieval returns the full model. Use the summary or property tools when only an overview or selected fields are needed. Creating or updating a dashboard requires the full dashboard JSON, while patching is intended for narrower changes.

Panel query execution and several datasource query families are disabled by default and need explicit enablement. Image rendering requires Grafana Image Renderer. Datasource operations may not work correctly on Grafana versions earlier than 9.0. Creating an incident can notify or alert many people, so the tool documentation requires confirmation and judicious use. Datasource creation must be performed in two calls: first obtain the field schema, then collect required values from the user before creating the datasource.