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

**Armature** (YC P26), founded by Theodore and Louis, launched on Show HN with product analytics and evals aimed at agent sessions that call an MCP server. The core claim is session reconstruction: from the MCP tool calls a server receives, Armature rebuilds the full interaction, including what the user asked the agent to do and what the agent thought while deciding on those calls.

Wrapping an existing MCP takes three lines of code through an SDK available in TypeScript, Python, and Go. Once instrumented, the product treats the tool-call stream as enough signal to reconstruct sessions in a dashboard. Operators see those sessions as readable conversations comparable to what a user would have run inside Claude or ChatGPT, plus a ranking of the MCP’s most popular use cases.

For MCP builders, tool logs alone rarely answer product questions. A single call sequence does not show the user goal, the agent’s plan, or which workflows drive real load. Session reconstruction and use-case ranking give server authors the same kind of funnel and feature-usage signal app teams expect from product analytics, applied to agent-mediated traffic rather than direct UI clicks.

The pitch sits in the MCP observability layer rather than competing as another agent host or chat client. As more products expose tools to Claude, ChatGPT, and similar agents, server teams need visibility into how agents actually use those tools. Armature’s bet is that thin SDK wrapping plus reconstructed sessions is enough to make that traffic inspectable without forcing a new runtime.

Practical next step for MCP authors is to instrument with the three-line SDK and check whether reconstructed sessions and use-case rankings match real production traffic. Watch how evals sit next to those sessions in the dashboard, and whether the wrap model stays simple once multiple tools, multi-turn agent loops, and production volume hit the same pipeline.