[bot] Missing instrumentation for the official Model Context Protocol (MCP) Java SDK

[braintrust-bot]

Summary

The Braintrust Java SDK does not instrument the official Model Context Protocol (MCP) Java SDK (io.modelcontextprotocol.sdk:mcp). MCP is the open standard for connecting LLM applications to external tools/data sources, and its Java SDK is maintained by the Model Context Protocol org in collaboration with Spring AI — a framework this repo already instruments (springai_1_0_0). The MCP client's tool-invocation execution surface (the mechanism models actually use to call tools) is completely uninstrumented, even though Spring AI chat model spans are captured.

What is missing

The io.modelcontextprotocol.sdk:mcp SDK provides these execution-oriented surfaces, none of which are instrumented:

API	Class	Description
callTool() / callToolAsync()	McpSyncClient / McpAsyncClient	Executes a tool on an MCP server and returns its result — the core "tool-calling during model execution" surface
listTools()	McpSyncClient / McpAsyncClient	Discovers available tools from a server
getPrompt() / listPrompts()	McpSyncClient	Retrieves/executes server-defined prompt templates
readResource() / listResources()	McpSyncClient	Reads server-exposed resources referenced during agent execution
Server-side CallToolResult handlers	McpServer (sync/async)	Server-side tool execution when this JVM acts as an MCP server

Because Spring AI's MCP integration (spring-ai-mcp) wraps this same SDK to expose MCP tools as ToolCallbacks to ChatModel, any tool call made through MCP during an instrumented Spring AI chat completion still produces no dedicated tool-execution span — the model call is traced, but the tool invocation it triggers is not.

Why existing instrumentation does not cover this

• springai_1_0_0 only instruments OpenAiChatModel$Builder and AnthropicChatModel$Builder (see issue [BOT ISSUE] Spring AI instrumentation limited to OpenAI and Anthropic chat models; no embedding model support #71 for other gaps in that module) — it has no handler for MCP tool-callback invocation.
• langchain_1_8_0 instruments LangChain4j's own TracingToolExecutor, which is unrelated to and does not intercept MCP client calls.
• No other module references MCP classes at all.

Braintrust docs status

• Braintrust's own docs page for MCP (https://www.braintrust.dev/docs/integrations/developer-tools/mcp) describes Braintrust's MCP server for AI coding tools (Claude Code, Cursor, etc.) — a completely different concern from instrumenting a user's own MCP client/server code in a Java application: not_found for MCP client/server execution instrumentation
• No reference to the MCP Java SDK appears in the Java-specific tracing docs at https://www.braintrust.dev/docs/instrument/trace-llm-calls

Upstream sources

• Official Java SDK repository: https://github.com/modelcontextprotocol/java-sdk — "The official Java SDK for Model Context Protocol servers and clients. Maintained in collaboration with Spring AI"
• Maven Central: https://central.sonatype.com/artifact/io.modelcontextprotocol.sdk/mcp — latest release v2.0.0
• MCP Java SDK docs: https://modelcontextprotocol.info/docs/sdk/java/

Local files inspected

• settings.gradle (lines 28–33) — lists all instrumentation modules (openai_2_15_0, anthropic_2_2_0, genai_1_18_0, langchain_1_8_0, springai_1_0_0, aws_bedrock_2_30_0); no MCP module present
• braintrust-sdk/instrumentation/springai_1_0_0/src/main/java/dev/braintrust/instrumentation/springai/v1_0_0/BraintrustSpringAI.java — only handles OpenAiChatModel$Builder / AnthropicChatModel$Builder; no MCP tool-callback handling
• Full-repo grep for modelcontextprotocol, io.modelcontextprotocol — zero matches