MCP servers

The MCP servers page in AI Observability lets you monitor MCP Server health, traffic, and tool execution across all published endpoints. Navigate to Monitor > AI Observability > MCP servers to open the page.

MCP request data is captured automatically. No configuration is required to enable audit logging.

Snaplex nodes aggregate MCP data every minute and the data appears on the MCP servers page one or two minutes later. Use the time range selector at the top of the page to view data for 1h, 24h, 7d, or 30d, or select a custom date range.

Non-admin users can view data on assets for which they have at least read permission. Environment admins can view details about all assets in the environment (Org). All data is scoped to the current Org.


MCP servers page in Monitor showing the fleet overview dashboard and server list

Dashboard cards

Three cards at the top of the page summarize MCP traffic across all servers in the current time range:

  • TRAFFIC: A sparkline showing the number of MCP requests per minute.
  • ERRORS: A sparkline showing the server error rate as a percentage of total requests.
  • LATENCY P99: A bar chart showing the 99th percentile request latency in milliseconds across all MCP calls.

MCP Server list

The server list shows one row per MCP Server, sorted by calls (highest first) by default. Use the filter chips and search bar to narrow the list.

Column Description
MCP Server The server name, the number of tools it exposes, and its current requests-per-minute rate.
Calls The total number of MCP requests received by the server during the selected time range.
Error rate The percentage of MCP requests that resulted in an error during the selected time range.
Latency p99 The 99th percentile request latency for the server during the selected time range.
Last called The timestamp of the most recent MCP request received by the server.

Server detail

Click a row in the server list to open the server detail view. The header displays the server name and the following action links:

  • Edit in Designer: Opens the server's pipeline in Designer.
  • View execution: Opens the execution history for the server's pipeline in Monitor.
  • Download logs: Downloads the request log for the server.

A summary metadata grid below the header shows the following fields:

  • Path: The project path of the MCP Server asset.
  • Server URL: The endpoint URL for the MCP Server.
  • Server ID: The unique identifier for the MCP Server.
  • Errors: The count of errors in the selected time range.
  • Last called: The timestamp of the most recent request.
  • Requests/min: The current request rate.
  • Calls: The total request count in the selected time range.
  • Auth: The authentication method configured for the server.
  • Tools count: The number of tools exposed by the server.
  • Error rate: The error rate as a percentage.
  • Latency p99: The 99th percentile latency.

The detail view contains two tabs: Tools and Requests.


MCP Server detail view showing the Tools tab with per-tool metrics and expanded mini charts

Tools tab

The Tools tab lists each tool exposed by the selected MCP Server. Expand a tool row to see per-tool mini charts.

Column Description
Tool The name of the tool.
Calls The number of times the tool was called during the selected time range.
Error rate The percentage of calls to this tool that resulted in an error.
Latency p99 The 99th percentile latency for calls to this tool.
Last called The timestamp of the most recent call to this tool.

Expanding a tool row reveals three mini charts scoped to that tool: TRAFFIC, ERRORS, and LATENCY P99.

Requests tab

The Requests tab shows the MCP request log for the selected server. The log is filterable and paginated.


MCP Server detail view showing the Requests tab with the request log

Column Description
Method The MCP method name, such as tools/call.
Tool The name of the tool invoked by the request.
Time The timestamp when the request was received.
Auth The authentication method used for the request.
Client/Username The client identifier or the username used for authentication.
Remote Address The IP address of the client that sent the request.
Status The result status of the request.
Duration The end-to-end latency for the request.

Request detail

Click a row in the request log to open the request detail view. The following metadata fields are shown:

  • Tool: The tool that was invoked.
  • Method: The MCP method name.
  • Client: The client identifier.
  • Status: The result status of the request.
  • Remote Address: The client IP address.
  • Auth: The authentication method used.
  • Received: The timestamp when the request was received.
  • Latency: The end-to-end request latency.

Call payload

The Call payload section shows the request and response data:

  • REQUEST: The tools/call JSON sent by the client.
  • RESULT (on success): The JSON response returned by the pipeline.
  • ERROR (on failure): The error type and message, joined in a single block.

Body payloads larger than 2,048 bytes are truncated and appended with ...[truncated].

Pipeline execution

The Pipeline execution section is expandable and shows pipeline execution details for the pipeline that handled the MCP request.


Request detail view showing pipeline execution details for a successful request

When pipeline execution details are available, the following fields are shown:

  • Pipeline name and Path
  • Username: The user that executed the pipeline.
  • Snaplex and Node name
  • MCP task name: The Ultra Task that handled the request.
  • Prepared at, Executed at, and Completed at timestamps
  • Execution duration
  • Invoker
  • A Snap-level execution table with slot and per-Snap details

The section also includes Open in Designer and View execution links.


Request detail view showing the error state when pipeline execution details cannot be loaded

If pipeline execution details cannot be retrieved, the section displays Could not load pipeline execution.

Note: tools/list requests display the request and result JSON in the flyout but do not include a Pipeline execution section because they do not invoke a pipeline. initialize requests appear in the log and open a flyout, but the Request and Result fields are not populated with content.

Limitations

  • Body payloads over 2,048 bytes are truncated with ...[truncated].
  • The pipeline_run_id and correlation_id fields are populated only for Ultra Task execution paths. These fields may be blank for Triggered Task executions.
  • initialize requests open a flyout, but the Request and Result fields are not populated with content.
  • resources/read requests are not yet captured.