Confluent MCP: Kafka topics and consumer lag for agents

A stuck consumer, a topic backed up by hundreds of thousands of messages, and you’re bouncing between the CLI, a console and dashboards. Confluent MCP turns everyday Kafka operations into MCP tools, so an agent can list topics, check consumer-group lag and pull a few messages to see their shape. Confluent open-sourced it, and it had about 170 stars as of 2026-10-06, so it’s little known outside Kafka circles.

What it does

• Topics and messages: list-topics, create-topics and delete-topics manage topics; produce-message and consume-messages send and read.
• Consumer groups: list-consumer-groups, describe-consumer-group and get-consumer-group-lag show state and lag.
• Schema Registry: list-schemas, create-schema and delete-schema manage schemas.
• Confluent Cloud extras: with cloud credentials it adds Flink SQL statements, Flink catalogs, connectors, Tableflow and more, over 50 tools in total per the README.
• Self-diagnosis: explain-disabled-tools, list-configured-connections and config-help need no credentials and show which tools are disabled for missing config and what YAML to add.

Who it’s for

• Backend and platform engineers who want an agent to track down backlogs and consumer lag.
• Data engineers writing Flink SQL on Confluent Cloud who’d like to submit and inspect statements in plain language.

Setup

Needs Node.js 22.19 or newer. Generate a config, fill in your Kafka endpoints (a local cluster needs only bootstrap_servers), then start it:

npx @confluentinc/mcp-confluent --init-config
npx @confluentinc/mcp-confluent --config ./config.yaml

Per-client setup guides are in the repo’s docs/ folder. For Confluent Cloud you can skip static keys entirely: the first call to a cloud-authenticated tool opens a browser for OAuth sign-in, and later calls reuse that session.

Our take

Streaming platforms rarely get a solid MCP server, and this one comes from the leading Kafka vendor without locking you to its cloud: plain self-managed Kafka and Schema Registry work, and so does a self-managed Confluent Platform cluster, with the same tool set as any local deployment. Confluent also runs a fully managed MCP server with nothing to run locally, worth taking instead if zero ops is the priority. Tools enable themselves from your config, and --allow-tools or --block-tools give precise control; block delete-topics and produce-message from the start if production data is in reach. Two things to know: the README says the open-source version gets best-effort community support, with a managed version for Confluent Cloud customers, and it sends usage data and errors to Sentry by default unless you set DO_NOT_TRACK=true. In the data category it doesn’t overlap with DuckDB / MotherDuck MCP: one handles message streams, the other analytical queries. Licensed MIT.