Anthropic ships MCP Observability API and hosted traces for agents
Anthropic rolled out first-party observability for Model Context Protocol workflows, giving teams production-grade tracing and metrics for agent stacks built on Claude and MCP tools.
Anthropic quietly expanded the Model Context Protocol ecosystem this week with a new MCP Observability API and a hosted trace explorer baked into the Claude console and docs. Any agent using MCP servers and tools can now emit standardized spans describing tool calls, retries, parallel branches, and sub-agent invocations. The system correlates those spans with model tokens, latency, and cost data, giving teams an end-to-end view of how real user requests propagate through their agent workflows.
What changed. Anthropic introduced first-party observability for MCP-based agents, including a schema for traces, SDK helpers to emit spans, and a hosted UI for inspecting tool call chains, errors, and performance.
Why it matters. As enterprises scale from a handful of experimental agents to dozens in production, lack of visibility into tool interactions and orchestration logic has been a core blocker—this turns MCP from a wiring standard into something much closer to a production runtime.
Builder takeaway. If you’ve been instrumenting Claude agents with ad hoc logs, migrate critical workflows to the MCP Observability primitives so you can debug prompt injection fallout, tool failures, and slow chains with structured traces instead of scattered console output.