Show HN: Product analytics (and evals) for agent sessions on your MCP

Theodore and Louis, founders of Armature (YC P26), launched product analytics and evals aimed at MCP tool providers. The product reconstructs the full agent session behind each MCP tool call a server receives, including what the user asked their agent to do and what the agent thought. Integration is a three-line wrap of an existing MCP, with an SDK in Typescript, Python, and Go.

On the technical side, the system sits at the MCP boundary rather than inside the host chat app. Once the MCP is wrapped, Armature rebuilds sessions from the tool-call stream so operators can read a reconstructed conversation comparable to the one the user had in Claude or ChatGPT. The dashboard also ranks the MCP’s most popular use cases from that same session data.

For engineers shipping MCP servers, tool-call logs alone rarely show intent or agent reasoning. Reconstructing the user prompt and agent thought trail turns opaque invocations into inspectable sessions, which matters for debugging failures, judging whether tools match real workflows, and running evals against how agents actually use the surface.

In market terms, MCP has become a common way to expose tools to agents in products like Claude and ChatGPT, while visibility into those sessions has stayed thin for server owners. Armature positions itself as analytics and evals for that layer: not another host client, but instrumentation for the people who build and operate the tools those clients call.

Practical next step for builders is to wrap a production or staging MCP with the three-line SDK and inspect reconstructed sessions plus the use-case ranking in the dashboard. Watch whether the reconstructed sessions surface unexpected tool sequences or dominant use cases that should drive API shape, docs, and eval suites.