sanjay-amu/kafka-sentinel-mcp

This project is a bridge between AI assistants and Kafka, the widely used data streaming platform. Kafka moves large volumes of messages between different parts of a software system, and keeping it healthy is a major concern for operations teams. The creator built this server so that an AI agent can look at a Kafka system, understand what is going on, and help diagnose problems, all without having the ability to change or break anything. The creator's background is in running Kafka for financial systems with extremely high uptime requirements, and he wanted a way for AI to help without needing dangerous administrative access. The core idea is that when a data pipeline fails at odd hours, the questions an engineer asks are repetitive. Is the problem a consumer falling behind, a stuck data partition, or a rebalance storm? These are pattern-matching tasks, which AI is good at. This server gives the AI a set of tools to answer these questions. It can list all topics and consumer groups, check cluster health, and measure consumer lag. It can also audit topic settings to find configurations that might violate good practices, check partition state, and determine if data can be replayed after a failure. One tool bundles all this information into a single snapshot for a postmortem report. A key feature is the strict security model. Every tool is read-only by design. The AI cannot produce messages, change settings, commit offsets, or alter access controls. This is enforced at the code level, where administrative operations are not even imported. The server supports secure connections, and credentials are passed via environment variables without being logged. Every action the AI takes is logged for audit purposes. The recommendation is to run the server with a Kafka user account that only has permission to describe and read data, not to change it. The project is in an early stage. It installs with a standard Python package manager and connects to a Kafka cluster using a simple configuration. You can then connect it to an AI client like Claude and ask it questions in plain English about your specific consumer groups and topics. The creator welcomes feedback, especially stories about what information would have been useful during a real incident. The project uses the MIT license.