LangSmith MCP

LangSmith MCP is the Model Context Protocol server for LangSmith, LangChain's observability and evaluation platform for LLM applications. It lets an assistant read the trace store and the prompt library directly, so debugging a production run does not require leaving the editor for the LangSmith web app. The server exposes tracing and prompt data as tools. An agent can list tracing projects, fetch runs and traces with filtering, and pull the full history of a conversation thread to see what a user and an application actually exchanged. On the prompt side it can list prompts, fetch a specific prompt by name, and push a prompt back to the hub. Datasets, experiment results and account usage are reachable as well. Responses use character budget pagination, so large traces come back in slices sized to fit a model's context instead of truncating unpredictably. The audience is engineering teams already instrumenting their agents or chains with LangSmith. The typical use is investigative: an eval regresses or a user reports a bad answer, and the assistant pulls the failing run, reads the intermediate steps and tool calls, then correlates them with the prompt version stored in the hub. It is mostly a read path onto data the LangSmith SDK already collects, which makes it a low risk addition next to write heavy connectors.