<!-- library-gap-audit: mcp-go-sdk-instrumentation -->
Summary
The official Model Context Protocol (MCP) Go SDK (github.com/modelcontextprotocol/go-sdk, package github.com/modelcontextprotocol/go-sdk/mcp) is not instrumented anywhere in this repo. MCP is the standard protocol for connecting LLM applications/agents to external tools and context; the Go SDK is maintained by the modelcontextprotocol org in collaboration with Google (the same governance model backing MCP's TypeScript/Python reference SDKs), and it defines the client/server execution surface agents use to discover and call tools during a run — squarely in scope as an agent-run/tool-calling execution surface, not a CRUD/admin API.
This repo already instruments other agent-execution frameworks separately from base model SDKs (e.g. Google ADK alongside google.golang.org/genai, LangChainGo alongside base providers), so a dedicated trace/contrib package for MCP client/server calls fits the existing pattern.
Weekly downloads
The Go SDK itself has no published download-count registry (it's a Go module without npm/PyPI distribution). As an adoption proxy, the official reference Python SDK for the same protocol:
Weekly downloads: 87,149,707 (as of 2026-08-03; https://pypistats.org/packages/mcp — package mcp, the official Python MCP SDK)
What is missing
A trace/contrib/mcp/ (or similarly named) integration module wrapping the execution-relevant calls on mcp.Client / mcp.Server:
Client.CallTool — the core tool-execution call agents make during a run (request/response, potential streaming progress notifications)Server tool-handler invocation — the server-side execution of a registered tool handler, relevant when this repo's users build MCP servers that Braintrust should traceClient.Session request/response lifecycle for tool listing and invocation (ListTools, CallTool) which represent the meaningful action surface (not resource/prompt CRUD)
Braintrust docs status
not_found. Braintrust's own MCP server (documented at https://www.braintrust.dev/docs/reference/mcp) is a product that exposes Braintrust's own data to coding assistants — it is not tracing/instrumentation of MCP client or server execution. The tracing integrations page (https://www.braintrust.dev/docs/guides/tracing/integrations) lists tracing integrations (OpenTelemetry, Vercel AI SDK, LangChain, LlamaIndex, DSPy, Instructor, LiteLLM, etc.) and agent frameworks (OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, AutoGen, Google ADK, Strands, etc.) — MCP does not appear in either list, in this repo or in any other Braintrust SDK's documented integrations.
Upstream sources
Braintrust docs sources checked
Local repo files inspected
go.mod — no modelcontextprotocol/go-sdk dependencytrace/contrib/ — no mcp/ directory existstrace/contrib/all/all.go — no MCP importexamples/internal/ — no MCP example subdirectory- Repo-wide search for
modelcontextprotocol, mcp-go, "mcp" — no matches anywhere in the codebase
<!-- library-gap-audit: mcp-go-sdk-instrumentation -->
Summary
The official Model Context Protocol (MCP) Go SDK (
github.com/modelcontextprotocol/go-sdk, packagegithub.com/modelcontextprotocol/go-sdk/mcp) is not instrumented anywhere in this repo. MCP is the standard protocol for connecting LLM applications/agents to external tools and context; the Go SDK is maintained by themodelcontextprotocolorg in collaboration with Google (the same governance model backing MCP's TypeScript/Python reference SDKs), and it defines the client/server execution surface agents use to discover and call tools during a run — squarely in scope as an agent-run/tool-calling execution surface, not a CRUD/admin API.This repo already instruments other agent-execution frameworks separately from base model SDKs (e.g. Google ADK alongside
google.golang.org/genai, LangChainGo alongside base providers), so a dedicatedtrace/contribpackage for MCP client/server calls fits the existing pattern.Weekly downloads
The Go SDK itself has no published download-count registry (it's a Go module without npm/PyPI distribution). As an adoption proxy, the official reference Python SDK for the same protocol:
Weekly downloads: 87,149,707 (as of 2026-08-03; https://pypistats.org/packages/mcp — package
mcp, the official Python MCP SDK)What is missing
A
trace/contrib/mcp/(or similarly named) integration module wrapping the execution-relevant calls onmcp.Client/mcp.Server:Client.CallTool— the core tool-execution call agents make during a run (request/response, potential streaming progress notifications)Servertool-handler invocation — the server-side execution of a registered tool handler, relevant when this repo's users build MCP servers that Braintrust should traceClient.Sessionrequest/response lifecycle for tool listing and invocation (ListTools,CallTool) which represent the meaningful action surface (not resource/prompt CRUD)Braintrust docs status
not_found. Braintrust's own MCP server (documented at https://www.braintrust.dev/docs/reference/mcp) is a product that exposes Braintrust's own data to coding assistants — it is not tracing/instrumentation of MCP client or server execution. The tracing integrations page (https://www.braintrust.dev/docs/guides/tracing/integrations) lists tracing integrations (OpenTelemetry, Vercel AI SDK, LangChain, LlamaIndex, DSPy, Instructor, LiteLLM, etc.) and agent frameworks (OpenAI Agents SDK, Claude Agent SDK, LangGraph, CrewAI, AutoGen, Google ADK, Strands, etc.) — MCP does not appear in either list, in this repo or in any other Braintrust SDK's documented integrations.
Upstream sources
Braintrust docs sources checked
Local repo files inspected
go.mod— nomodelcontextprotocol/go-sdkdependencytrace/contrib/— nomcp/directory existstrace/contrib/all/all.go— no MCP importexamples/internal/— no MCP example subdirectorymodelcontextprotocol,mcp-go,"mcp"— no matches anywhere in the codebase