Model Context Protocol

Rules

What the lists below cannot show: these rules live in the convention's prose, not in its registry. Paraphrased from main/docs/gen-ai/mcp.md; the upstream wording wins.

• SHOULDPropagate context in the message, not the transport

  Inject the trace context into the request's params._meta, using the unprefixed keys traceparent, tracestate and baggage (SEP-414). The receiver extracts it from there and uses it as the remote parent.

• SHOULDThe transport span is a link, not the parent

  One Streamable HTTP request can carry several MCP messages, and one message can span several HTTP requests on retry. The server span takes its parent from params._meta and links the ambient (HTTP) context if there is one.

• SHOULDName spans after the method and its target

  {mcp.method.name} {target}, where the target is gen_ai.tool.name or gen_ai.prompt.name: tools/call get-weather. With no low-cardinality target, just {mcp.method.name}. mcp.resource.uri goes in the name only when a user opts in.

• SHOULDA tool that returns isError is a tool_error

  A CallToolResult with isError: true is a successful JSON-RPC call, so it carries no error code. Set error.type to tool_error. A JSON-RPC error sets rpc.response.status_code and error.type to the code; an HTTP failure without one uses the HTTP status code.

• SHOULD NOTThe caller's mistakes are not server errors

  On the server, -32700, -32600, -32601, -32602 (including an unknown tool) and -32002 mean the caller sent something the server could not serve. They are not errors on the server span; every other code is.

• SHOULD NOTOne span per tool call

  If GenAI instrumentation already traces the tool execution, the MCP client adds its attributes to that execute_tool span instead of creating another. Set gen_ai.operation.name to execute_tool on tool calls only.

• SHOULDMCP conventions over RPC ones

  MCP is JSON-RPC, but the RPC and HTTP conventions miss the messages inside a streaming call. Use mcp.method.name rather than rpc.method.

• SHOULDDuration buckets from 10 ms to 5 minutes

  The operation and session duration histograms are in seconds, with explicit bucket boundaries [0.01, 0.02, 0.05, 0.1, 0.2, 0.5, 1, 2, 5, 10, 30, 60, 120, 300].

• MAYArguments and results are opt-in

  gen_ai.tool.call.arguments, gen_ai.tool.call.result and prompt arguments (gen_ai.prompt.variable.<name>) may hold sensitive data, so they are recorded only when a user opts in.

Signals

What this domain emits, and the attributes each one carries, most strongly required first. An attribute marked * is required on that signal; * is conditionally required.

Metrics 4

• mcp.client.operation.duration
  Developmenthistogram · sThe duration of the MCP request or notification as observed on the sender from the time it was sent until the response or ack is received.
• mcp.client.session.duration
  Developmenthistogram · sThe duration of the MCP session as observed on the MCP client.
• mcp.server.operation.duration
  Developmenthistogram · sMCP request or notification duration as observed on the receiver from the time it was received until the result or ack is sent.
• mcp.server.session.duration
  Developmenthistogram · sThe duration of the MCP session as observed on the MCP server.

Specification

• Spansare bound by the trace specification — 302 requirements across 7 documents, 168 of them MUST or SHALL.
• Metricsare bound by the metrics specification — 561 requirements across 8 documents, 287 of them MUST or SHALL.

The specification prose never names this domain. That is not an omission: the specification defines signals, not subject matter — what binds mcp telemetry is the general rules above, and the conventions themselves carry the rest.

On the wire, this lands in trace/v1/trace, metrics/v1/metrics. Field-by-field definitions are under OTLP.

Attributes 4

Recent changes

Breaking and notable changes to this domain across the last 6 semantic-conventions releases.