Toward an MCP Observability Specification

The Model Context Protocol defines how agents discover and invoke tools. It defines resources, prompts, and transport mechanisms. It standardizes the interface between an agent and the capabilities it can use. This is significant work, and it has enabled an ecosystem of interoperable MCP servers to emerge in a short time. But MCP does not define how agents should report what they did. There is no standard trace format. No standard eval interface. No standard way to express cost metadata, token usage, or span relationships. Every observability solution in the MCP ecosystem — including Iris — is bolted on after the fact. We build our own schemas, define our own tool interfaces, and store data in our own formats. The protocol that standardized tool invocation has nothing to say about tool observation. I think this is a gap worth closing. This post is a sketch of what an MCP observability specification could look like, grounded in what I have learned building Iris as an MCP-native observab