Edit

Monitor MCP server traffic in Azure API Management

In this article, you learn what telemetry Azure API Management emits for traffic to MCP servers, how to enable payload logging for tool arguments and results, and how to query the data in Azure Monitor.  

Prerequisites 

Default telemetry for MCP servers

For every MCP request, API Management writes an Application Insights requests row with MCP-specific dimensions and sets the standard duration field. You can chart per-tool latency without changing any configuration. For details, see the MCP telemetry reference section, later in this article.  

Note

MCP telemetry follows the OpenTelemetry semantic conventions for generative AI, which define standard telemetry attribute names (for example, gen_ai.*) so data is consistent across tools.

Enable payload logging for arguments and results 

By default, API Management doesn't capture the arguments and results of tool calls. To enable capture for an MCP server: 

  1. In the Azure portal, go to your API Management instance. 

  2. Select APIs > MCP servers, and then select the MCP server you want to log. 

  3. Select Settings > Diagnostic Logs. 

  4. Enable Frontend and Backend payload logging. Select Save. 

Caution

Tool arguments and results can include prompts, customer data, or secrets. Enable payload logging only for the MCP servers and environments where you need it. Apply scrubbing or claim allowlists before broad rollout. 

Query MCP traffic with KQL 

The following are sample Kusto queries you can run in Azure Monitor to analyze MCP traffic. In these examples, replace sales-mcp with the name of your MCP server where applicable.

List the last 50 tool calls on a given MCP server

requests
| where customDimensions["api.type"] == "Mcp"
  and customDimensions["service.name"] == "sales-mcp"
  and customDimensions["gen_ai.operation.name"] == "tools/call"
| project timestamp,
          tool       = customDimensions["gen_ai.tool.name"],
          session    = customDimensions["gen_ai.conversation.id"],
          client     = strcat(customDimensions["user_agent.name"], "/",
                              customDimensions["user_agent.version"]),
          durationMs = duration,
          success
| order by timestamp desc
| take 50

Top MCP clients by tool-call volume 

requests
| where customDimensions["api.type"] == "Mcp"
  and customDimensions["gen_ai.operation.name"] == "tools/call"
| summarize calls = count()
    by client = strcat(customDimensions["user_agent.name"], "/",
                       customDimensions["user_agent.version"])
| top 10 by calls desc

p50 and p95 latency per tool over the last 24 hours

requests
| where customDimensions["api.type"] == "Mcp"
  and customDimensions["gen_ai.operation.name"] == "tools/call"
  and timestamp > ago(24h)
| summarize p50   = percentile(duration, 50),
            p95   = percentile(duration, 95),
            calls = count()
    by tool = tostring(customDimensions["gen_ai.tool.name"])
| order by p95 desc

Per-tool error rate over time

requests
| where customDimensions["api.type"] == "Mcp"
  and customDimensions["gen_ai.operation.name"] == "tools/call"
| summarize total    = count(),
            failures = countif(success == false)
    by bin(timestamp, 5m),
       tool = tostring(customDimensions["gen_ai.tool.name"])
| extend errorRate = todouble(failures) / total
| render timechart

Inspect arguments sent to a specific tool

For this scenario, make sure payload logging is enabled for the MCP server.

requests
| where customDimensions["api.type"] == "Mcp"
  and customDimensions["service.name"] == "sales-mcp"
  and customDimensions["gen_ai.tool.name"] == "create_quote"
  and timestamp > ago(1h)
| project timestamp,
          session = customDimensions["gen_ai.conversation.id"],
          args    = customDimensions["gen_ai.tool.call.arguments"],
          result  = customDimensions["gen_ai.tool.call.result"]

Add custom dimensions with the trace policy 

To capture data that's not in the built-in schema - for example, a custom x-agent-id header, a JWT claim, or a correlation ID - use the trace policy at the scope of the MCP server. 

Warning

Don't access context.Response.Body from policies attached to MCP scope. MCP responses are streamed, and reading the body breaks the stream. 

MCP telemetry reference 

The following dimensions appear on every MCP request: 

Property Description
gen_ai.operation.name JSON-RPC method (tools/list or tools/call).
gen_ai.conversation.id MCP session ID.
network.protocol.name Protocol name (MCP).
network.protocol.version Protocol version.
auth.type Inbound authentication method.
user_agent.name MCP client name (for example, vscode or claude-desktop).
user_agent.version MCP client version.
service.name MCP server name.
service.version MCP server version.
api.type API type discriminator (Mcp).
error.message Error string, on failure.
error.type Error category, on failure.

Additional fields on tools/list

Metric Description
ToolCount Number of tools returned in the response.

Additional fields on tools/call

Property Description
gen_ai.tool.name Tool that the agent invoked.
gen_ai.tool.type Tool type.
gen_ai.tool.call.arguments Arguments JSON. Only present when payload logging is enabled.
gen_ai.tool.call.result Result JSON. Only present when payload logging is enabled.

Related content 

  • Last updated on