Support for OpenTelemetry Context Propagation for MCP via _meta field

[punitmahes]

Description

I am looking to implement end-to-end distributed tracing for an AI agent architecture using the Model Context Protocol (MCP).

Currently, the MCP Auto Instrumentation SDK propagates OpenTelemetry context via a reserved _meta field within the JSON-RPC message body, rather than standard HTTP headers. As the AI MCP Proxy plugin acts as a protocol bridge, I am proposing a feature to extract tracing context from this _meta field to ensure trace continuity through the Kong Gateway.

Background: Why _meta?

The Model Context Protocol (MCP) standardized the use of the _meta property bag for several reasons:

• Transport Agnosticism: MCP is designed to work over stdio, WebSockets, and SSE. Since stdio lacks HTTP headers, transport-independent propagation is required.
• Multiplexing: Over Streamable HTTP (SSE), multiple tool calls can occur within one long-lived HTTP request. Header-based tracing only captures the connection, whereas _meta allows each tool call to have its own unique trace/span parentage.
• Specification: This convention is formalized in the MCP specification (referencing PR #414), reserving keys like traceparent, tracestate, and baggage within the _meta object.

Proposed Capability

Since the AI MCP Proxy plugin already parses the request body to perform protocol translation, it would be highly beneficial if the plugin could:

1. Extract the traceparent (and tracestate/baggage) from $.params._meta.
2. Map this to the gateway's internal OpenTelemetry span as a remote parent.
3. Inject the updated context into the outgoing headers for the upstream service.

Example Payload

{
  "jsonrpc": "2.0",
  "id": "1",
  "method": "tools/call",
  "params": {
    "name": "get_weather",
    "arguments": { "location": "San Francisco" },
    "_meta": {
      "traceparent": "00-0af7651916cd43dd8448eb211c80319c-00f067aa0ba902b7-01",
      "tracestate": "...",
      "baggage": "..."
    }
  }
}

Questions / Discussion

• Existing Hooks: Does the current AI MCP Proxy plugin (or the core OpenTelemetry plugin) have an existing hook or configuration option to extract OTel context from the request body rather than headers?
• Roadmap: Is body-based context extraction on the formal roadmap for Kong’s AI Gateway observability features, especially given the rising adoption of the MCP standard?

References

• MCP Specification: Basic _meta
• OpenTelemetry Semantic Conventions for GenAI MCP

Happy to implment a solution if needed.

[tzssangglass]

hi @fffonion @spacewander PTAL