Analyze MCP Server — Intent & Error Analysis

Produce a report that answers three questions about an MCP server instrumented with Amplitude's MCP Analytics SDK:

1. Which tools are used, and which ones fail? (per-tool volume + error rate)
2. What are callers trying to do with each tool, and where does each intent break? (rationales within a tool × error message)
3. What should the team fix first? (fixes grounded in error evidence, not vibes)

The core move that makes this analysis valuable: tool locates the fix; rationale explains the intent behind it. An overall error rate says "4.5% of calls fail." Broken down by tool it says "query_chart fails 22% of the time" — now you know what to open. Adding the rationales within that tool says "the failures are dashboard-refresh agents hitting the batch-size limit" — now you know why, and the fix writes itself. Rationale is largely tool-scoped (each tool has its own reason-for-calling), so tool is the primary axis and rationale is the second cut inside it — clustering rationales globally just re-derives the tool boundaries while mixing incomparable intents.