Azure Managed Grafana Observability Skill

You have access to the Azure Managed Grafana MCP server (AMG-MCP) tools for querying observability data, dashboards, and connected data sources.

Available Tools

Workflow

General Investigation

1. Understand the scope — Identify which Azure resources, services, or data sources are involved
2. Discover data sources — Use amgmcp_datasource_list to see what's connected
3. Check subscriptions — Use amgmcp_query_azure_subscriptions to list accessible subscriptions
4. Query relevant data — Use the appropriate tool for the data type (metrics, logs, traces, KQL, SQL)
5. Search dashboards — Use amgmcp_dashboard_search to find existing dashboards with relevant visualizations
6. Analyze and recommend — Summarize findings with actionable next steps

Failure Investigation

1. Get failure overview — Use amgmcp_insights_get_failures to retrieve failure summary (failed requests, dependencies, exceptions)
2. Trace failures — Use amgmcp_query_application_insights_trace to follow request traces across services
3. Check metrics — Use amgmcp_query_resource_metric to correlate with resource-level metrics (CPU, memory, latency)
4. Query logs — Use amgmcp_query_resource_log for detailed log entries around the failure time
5. Check resource health — Use amgmcp_query_resource_graph to verify resource configuration and status

Azure Monitor Metrics Analysis

1. Discover metrics — Use amgmcp_query_resource_metric_definition to list available metrics for a resource
2. Query metric values — Use amgmcp_query_resource_metric with the desired metric name, aggregation, and time range
3. Correlate — Check related resources' metrics to identify cascading issues

Azure Data Explorer (Kusto) Queries

1. Get cluster metadata — Use amgmcp_kusto_get_metadata to list clusters, databases, and table schemas
2. Write and run KQL — Use amgmcp_kusto_query to execute Kusto Query Language queries

Common KQL patterns:

// Recent errors from a table
TableName
| where Timestamp > ago(1h)
| where Level == "Error"
| project Timestamp, Message, Source
| order by Timestamp desc
| take 100

// Aggregated error counts
TableName
| where Timestamp > ago(24h)
| summarize ErrorCount = count() by bin(Timestamp, 1h), Source
| order by Timestamp desc

// Percentile latency analysis
RequestTable
| where Timestamp > ago(1h)
| summarize p50 = percentile(Duration, 50),
            p95 = percentile(Duration, 95),
            p99 = percentile(Duration, 99)
  by bin(Timestamp, 5m)

SQL Server Queries

1. Get schema metadata — Use amgmcp_mssql_get_metadata to list databases, tables, and columns
2. Run queries — Use amgmcp_mssql_query to execute SQL queries

GenAI Agent Telemetry

Use amgmcp_insights_get_agents to analyze GenAI agent performance. This tool queries data following OpenTelemetry Generative AI Semantic Conventions and returns:

• Agent invocation counts
• Token usage (input/output)
• Latency distributions
• Error rates

Azure Resource Graph Queries

Use amgmcp_query_resource_graph to query resource inventory across subscriptions:

// List all VMs and their status
Resources
| where type == "microsoft.compute/virtualmachines"
| project name, location, properties.hardwareProfile.vmSize, powerState = properties.extended.instanceView.powerState.code

// Find resources by tag
Resources
| where tags["environment"] == "production"
| summarize count() by type

// Check NSG rules
Resources
| where type == "microsoft.network/networksecuritygroups"
| mv-expand rules = properties.securityRules
| project name, ruleName = rules.name, access = rules.properties.access, direction = rules.properties.direction

Best Practices

• Start with data source discovery — Always use amgmcp_datasource_list to understand what's connected before querying
• Check metric definitions first — Use amgmcp_query_resource_metric_definition before querying metrics to get correct names
• Get schema before querying — Use amgmcp_kusto_get_metadata or amgmcp_mssql_get_metadata to discover table/column names
• Use dashboards for context — Search for existing dashboards with amgmcp_dashboard_search — they often reveal which metrics and queries are most relevant
• Scope time ranges — Always specify reasonable time ranges to avoid excessive data retrieval
• Aggregate across instances — Use amgmcp_query_application_insights_trace when traces span multiple Application Insights instances

Troubleshooting