feat(mcp): add Braintrust tracing for the official MCP Go SDK

[Atharva Mhaske (atharvamhaske)]

Summary

Adds a new trace/contrib/mcp integration for
github.com/modelcontextprotocol/go-sdk (v1.7.0),
covering the MCP tool execution surface that agents use during a run.

• InstrumentClient — traces ClientSession.ListTools and ClientSession.CallTool, including in-flight notifications/progress during tool calls
• InstrumentServer — traces incoming tools/list and tools/call, with a separate handler child span (mcp.tools.handler [name]) under the RPC span
• Orchestrion — auto-instruments mcp.NewClient / mcp.NewServer
• Examples — examples/mcp and examples/internal/mcp (in-memory transport; no provider API keys)
• Tests — in-memory JSON-RPC (no VCR/API keys); orchestrion coverage via TestMCP

Why

The official MCP Go SDK had no Braintrust instrumentation. MCP is the standard protocol for agent tool discovery and execution; this matches existing patterns (e.g. ADK, LangChainGo) as a dedicated trace/contrib module.

Test plan

• go test -C trace/contrib/mcp ./...
• make ci (lint, mod-verify, test, build)
• CI green on PR
  No new VCR cassettes or provider API keys required for MCP tests.

Notes

• Client spans set metadata.api to ClientSession.CallTool / ClientSession.ListTools
• Progress notifications are recorded as span events and in braintrust.output_json.progress
• Registered in scripts/nested_modules.txt, go.work, and trace/contrib/all
• I haven't check command make examples as i don't have braintrust api key to do so(will have to check before merging)

Closes #190

[Abhijeet Prasad (AbhiPrasad)]

please try this out in braintrust and share a screenshot of the braintrust trace! I will review after that.

[Atharva Mhaske (atharvamhaske)]

please try this out in braintrust and share a screenshot of the braintrust trace! I will review after that.

sure but do i need api key or is it there included in free accounts?

[Abhijeet Prasad (AbhiPrasad)]

you should be able to use a free braintrust account! Try out https://www.braintrust.dev/docs/tracing-quickstart

[Atharva Mhaske (atharvamhaske)]

you should be able to use a free braintrust account! Try out https://www.braintrust.dev/docs/tracing-quickstart

Tried this out with a real Braintrust trace like you asked. Found a bug while checking it: client spans (ListTools/CallTool) and server spans (tools/list, tools/call, handler) were landing in separate traces with no shared root. Turns out the MCP transport wasn't passing trace context from client to server, so the server side just started its own fresh trace instead of continuing the client's.

Fixed it by injecting a W3C traceparent into the request's _meta field on the client side, and reading it back out on the server side. Used our own propagation.TraceContext{} instead of the global otel propagator, since that one defaults to a no-op unless an app sets it up separately.

Added a test that checks the client span, server span, and handler span all share the same trace ID, and that the server span is a child of the client's call span.

Verified against a real trace, screenshots below in comment. Everything now nests correctly under one trace.

cc: Abhijeet Prasad (@AbhiPrasad)

[Atharva Mhaske (atharvamhaske)]

Screenshot 1

Screenshot 2

Now it works as expected, if anything doesn't work on your end feel free to drop me, happy to fix it.

[Atharva Mhaske (atharvamhaske)]

Abhijeet Prasad (@AbhiPrasad) previous ci failed due to missing imports(i have fixed it), can it be triggered again please?

[Atharva Mhaske (atharvamhaske)]

small suggestion about CI. right now every push on a fork PR needs manual approval to run.

as now i have one merged PR already (#225). gitHub has a setting that can fix this: change "require approval for all outside collaborators" to "require approval for first-time contributors".
with this setting, a known contributor does not need approval after the first PR. New contributors still need approval, so security stays the same.

this can save review time on future PRs. we can have it cuz it didnt flag my ci failure at first point when i pushed and it did when you ran the workflow, so i could catch it early.

just a suggestion, happy either ways with repo rules :)

[Atharva Mhaske (atharvamhaske)]

Abhijeet Prasad (Abhijeet Prasad (@AbhiPrasad)) previous ci failed due to missing imports(i have fixed it), can it be triggered again please?

re run of ci needed Abhijeet Prasad (@AbhiPrasad)

[Atharva Mhaske (atharvamhaske)]

Abhijeet Prasad (@AbhiPrasad) can you ptal !!

[Abhijeet Prasad (AbhiPrasad)]

I think this might take some time to merge, because we aren’t sure if we want to support this in the near term. We don’t have mcp instrumentation in any of the other sdks yet.

[Atharva Mhaske (atharvamhaske)]

I think this might take some time to merge, because we aren’t sure if we want to support this in the near term. We don’t have mcp instrumentation in any of the other sdks yet.

sure no issues

[Abhijeet Prasad (AbhiPrasad)]

Sorry for the trouble Atharva Mhaske (@atharvamhaske) but I'm going to close this. I don't think we want to pursue mcp instrumentation at the current moment. but we might revisit in a couple of months.