How can we monitor daily MCP server invocation counts across agents?

Copilot Studio can report daily MCP-backed tool executions across agents through environment-level Application Insights telemetry, but it does not currently expose a separate MCP-server invocation counter or documented server-endpoint dimension.

Configure centralized telemetry

Environment-level telemetry is the appropriate scope because it centralizes agent and tool execution across a Copilot Studio environment. It is currently a preview feature and requires:

• A managed Power Platform environment.
• Power Platform administrator or Dynamics 365 administrator privileges.
• An administrator role in the Dataverse environment.
• An Application Insights resource and the prerequisites for exporting data to it. ^1^

Configure it as follows:

1. In Power Platform admin center, create an Application Insights export package.
2. Set Export type to Copilot Studio, configure the destination Application Insights resource, and save the export.
3. Run a test conversation that invokes an MCP tool.
4. Allow up to 24 hours for a new configuration to deliver telemetry.
5. In Azure portal → Application Insights → Logs, verify:
   • InvokeAgent and ExecuteTool rows appear in dependencies.
   • Child ExecuteTool rows share an operation_Id with their InvokeAgent row. ^2^

Before filtering MCP calls, discover the values actually emitted in your environment because the preview schema can evolve:

dependencies
| where timestamp > ago(7d)
| mv-expand Key = bag_keys(customDimensions) to typeof(string)
| summarize Events=make_set(name),
            SampleValue=take_any(tostring(customDimensions[Key])) by Key
| order by Key asc

Query daily MCP tool calls by agent

After confirming that gen_ai.tool.type identifies MCP in your data, run:

dependencies
| where timestamp >= ago(90d)
| where name == "ExecuteTool"
| extend AgentName = tostring(customDimensions["gen_ai.agent.name"]),
         AgentId   = tostring(customDimensions["gen_ai.agent.id"]),
         ToolName  = tostring(customDimensions["gen_ai.tool.name"]),
         ToolType  = tostring(customDimensions["gen_ai.tool.type"])
| where ToolType contains "MCP"
| summarize TotalCalls=count(),
            SuccessfulCalls=countif(success == true),
            FailedCalls=countif(success != true)
    by Day=bin(timestamp, 1d), AgentName, AgentId, ToolName, ToolType
| order by Day desc, AgentName asc, TotalCalls desc

This supplies:

• Daily call totals
• Agent name and ID
• MCP tool name/type
• Successful and failed executions
• Historical trends over the selected retention period

The telemetry schema records gen_ai.tool.name, gen_ai.tool.type, call ID, arguments, and result for each ExecuteTool span. The documented fields do not include a distinct MCP server name or endpoint. ^2^

Server-level totals and endpoint identity

Maintain a controlled mapping of ToolName → MCP server → endpoint, then join that mapping into the query. For example:

let McpMap = datatable(ToolName:string, ServerName:string, Endpoint:string)
[
  "server-prefix:tool-a", "Finance MCP", "https://mcp.contoso.com",
  "server-prefix:tool-b", "Finance MCP", "https://mcp.contoso.com"
];
dependencies
| where name == "ExecuteTool"
| extend ToolName=tostring(customDimensions["gen_ai.tool.name"]),
         Agent=tostring(customDimensions["gen_ai.agent.name"])
| join kind=leftouter McpMap on ToolName
| summarize Calls=count(), Success=countif(success == true),
            Failures=countif(success != true)
    by bin(timestamp, 1d), ServerName, Endpoint, Agent

This counts MCP-backed ExecuteTool spans—not every network request received by the server. Discovery calls, retries, authentication attempts, and connection failures might not correspond one-for-one with those spans. For authoritative HTTP request counts, endpoint-level failures, and retry traffic, instrument the MCP server or its API gateway and correlate requests using available call/trace identifiers.

Dashboard and API options

• Agent Insights Hub provides tool call totals, success/failure counts, success rate, response time, and daily tool trends, with scheduled daily synchronization and historical backfill up to 365 days. It remains tool-level rather than a documented MCP-server/endpoint inventory. ^3^
• For programmatic extraction, query Application Insights logs through Azure Monitor query client libraries or the Application Insights REST API. ^4^
• Application Insights log retention is configurable through the associated Log Analytics workspace; supported log retention ranges from 30 to 730 days. ^5^

Therefore, MCP calls can be distinguished from non-MCP tools through gen_ai.tool.type, subject to validating its emitted values, but Application Insights represents them as individual tool executions, not as a separate server-call entity.

References

1. Telemetry with Application Insights overview
2. Environment-level telemetry with Application Insights (preview)
3. Monitor agent performance with Agent Insights Hub in Copilot Agent Kit
4. Metrics in Application Insights
5. Azure Monitor service limits