manisha.dev

Lab build, 2026

An agent that reads the logs first

An MCP server that lets Claude query Application Insights and Log Analytics, propose a root cause for an incident, and show the queries behind every claim.

The idea

Most of my career has included production support: an alert fires, someone opens the logs, and the first twenty minutes go to finding the right query. An agent should do those twenty minutes and hand a person a hypothesis with evidence.

The design

A Model Context Protocol server exposes a small, read-only toolset over Azure Monitor: run a KQL query against Log Analytics, fetch failed requests and dependencies from Application Insights, pull the last deployment from the release history. A Claude Agent SDK agent uses those tools to walk from the alert to a candidate root cause.

The rule is that every claim carries its receipt. The agent’s answer cites the exact KQL it ran and the rows that support it, so the on-call engineer can rerun the query instead of trusting the summary.

Guardrails and evals

The tools are read-only by design, scoped by a managed identity with reader rights and nothing else. A promptfoo suite replays past incidents with known causes and scores whether the agent found them, how many queries it needed and whether its citations hold. Every run is traced, so a wrong answer can be debugged like any other bug.

This is a lab build: a working design I am developing in the open, not a client system.