Investigate MCP calls using Trace

You can configure your agentic applications that call remote Google Cloud MCP servers to enable those servers to generate spans that record the status of the calls. You can then use Cloud Trace, which stores and analyzes spans that record tool calls, method invocations, and prompts and responses, to help you understand the behavior of your agentic applications.

A span records a single function or operation that is part of a larger end-to-end operation. A trace represents a single end-to-end operation and is composed of spans. Therefore, a trace provides information about the sequence of functions called and overall latency, while a span provides status and latency information for a specific function.

About tools

Tools let an agent gather information outside of its training data, typically by interacting with external APIs. With Google Cloud MCP servers, your agentic applications can call tools that invoke Google Cloud API methods. For example, your application can call the list_log_entries tool to fetch recent log data.

The Model Context Protocol (MCP) is a specification that defines a standardized way for AI agents to communicate and interact with external tools, data sources, and resources. For a complete list of products that provide an MCP server and links to their reference documentation, see Supported products.

How trace context is passed to an MCP server

MCP reference pages, such as the documentation for list_log_entries, provide an example curl command that you can use to send HTTP requests to the tool. The JSON payload sent in the request is similar to the following:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "NAME",
    "arguments": {
      // provide these details according to the tool's MCP specification
    }
  },
  "id": 1
}

The value of the name field lists the tool name, such as list_log_entries. The arguments field contains the information required by the operation to complete the request.

The MCP standard defines a _meta field, which lets clients and servers attach metadata to their interactions. Clients and servers can use this field to pass the trace context:

{
  "jsonrpc": "2.0",
  "method": "tools/call",
  "params": {
    "name": "NAME",
    "arguments": {
      // provide these details according to the tool's MCP specification
    },
    "_meta": {
        "traceparent": "00-TRACE_ID-PARENT_SPAN_ID-SAMPLED_FLAG",
        "tracestate": "Vendor specific information."
      },
  },
  "id": 1
}

The _meta data includes a traceparent field. The format of the value of this field is defined by the World Wide Web Consortium (W3C) traceparent header. The value includes the following:

• Version of the traceparent specification (00).
• ID of the trace (TRACE_ID).
• ID of the calling span (PARENT_SPAN_ID).
• Whether the calling span sampled the request (SAMPLED_FLAG). This field has a value of 01 when sampled and 00 when not sampled.

The tracestate field carries vendor-specific tracing information.

Limitations

• The trace context must follow the W3C Trace Context standard, and the sampled flag must be set to 1.

• Remote Google Cloud MCP servers can generate a single span for a tools/call operation, but they don't generate spans for other types of operations or child spans for the tools/call operation.

• Remote Google Cloud MCP servers generate a span only when the request is authenticated, authorized, and passes other internal checks.

MCP servers that support tracing

This section lists the remote Google and Google Cloud MCP servers that support tracing. It also describes how your self-hosted MCP servers can generate trace spans.

Self-hosted servers

We recommend that you instrument your self-hosted MCP servers to send telemetry data to your Google Cloud project. If you use Python, then the instrumentation can generate spans for tools/call operations. For more information, see Instrument a self-hosted MCP server with OpenTelemetry.

Remote Google and Google Cloud MCP servers that support tracing

The following products contain an integration that lets their remote MCP server generate a trace span for a tools/call operation: