Confluent MCP Server

The Confluent MCP Server connects AI assistants such as Claude Code or Cursor to Apache Kafka and Confluent's data streaming platform. Instead of looking up command-line commands or clicking through the web interface, you phrase a task, for example "List all topics in the cluster" or "Why did my connector fail?", and the assistant calls the matching tools. According to the repository the project belongs to the confluentinc organization, is released under the MIT license, currently has around 170 stars, and was last updated on October 21, 2026; the npm package is at version 1.6.0. An important caveat: the README explicitly describes the open-source server as a community-supported project for which Confluent provides no dedicated support and no service-level commitments.

What can the Confluent MCP Server do?

According to the README the server offers more than 50 tools. They include Kafka tasks such as creating, listing, and deleting topics, producing and consuming messages, describing consumer groups, and computing their lag. Added to that are Flink SQL (creating statements, fetching results, checking the health and errors of a statement), the Schema Registry for data schemas, Kafka Connect connectors with status, error summaries, and restarts, tags in the data catalog, Tableflow, metrics, and billing costs. There is also a search of the Confluent documentation and diagnostic tools that explain why a tool is currently unavailable.

Cloud, Confluent Platform, or local

The server enables tools based on your configuration. With Confluent Cloud all areas are usable. Against a local Kafka cluster or a self-managed Confluent Platform only Kafka and Schema Registry tools are available; Flink, Tableflow, metrics, and billing stay disabled there. That makes the server interesting for local development with a Docker cluster too, without any cloud account.

Setup and sign-in

The prerequisite is Node.js 22.19.0 or later. With npx @confluentinc/mcp-confluent --init-config you generate a config.yaml, enter your connection, and start the server with --config. Against Confluent Cloud, sign-in via OAuth in the browser is possible so no API keys need to be created; however, according to the README this so far works only for some of the tool areas, while others still need static API keys. Available transports are stdio, HTTP, and SSE. With the options --allow-tools and --block-tools you limit the tools deliberately, for example to read-only use.

Managed server from Confluent

Confluent additionally runs a fully managed MCP server for Confluent Cloud which, according to the documentation, needs no local installation and respects your existing roles and permissions (RBAC). Most of its tools are read-only; only actions such as restarting connectors or creating Flink statements change resources. The open-source server described here is the choice when you run self-managed Kafka or want to adjust the scope of tools.

What is it particularly useful for?

Its value lies in diagnosis and orientation: which topics exist, where is a consumer group backing up, why does a connector fail, what does the schema look like? Especially when troubleshooting, the assistant saves you from gathering commands. Compared with the command line it delivers answers in prose and can chain steps.

Security and privacy

If you grant write permissions you also allow the assistant to delete topics or schemas and to send messages. Use keys that are as narrow as possible, block destructive tools with --block-tools, and review actions before confirming. According to the README the server collects usage data and reports runtime errors to Sentry; both can be switched off with DO_NOT_TRACK=true. For HTTP and SSE you should not disable authentication.

Costs

The server itself is free to use. Costs arise for Confluent Cloud itself. The pricing page names the cluster types Basic, Standard, Enterprise, Freight, and Dedicated, bills by usage, and knows annual commitments with discounts; according to the page Basic has a free starting point. Current terms are on the official pricing page.

Who is the Confluent MCP Server for?

It is worthwhile for teams that run Kafka or Confluent Cloud and want to handle administration and troubleshooting in conversation. Without a Kafka environment it is useless. Try it first against a local test cluster or with read-only tasks.

Node.js 22.19.0 or later, an MCP client such as Claude Code or Cursor, and a Kafka environment: a local cluster, a Confluent Platform, or a Confluent Cloud account with API keys or OAuth sign-in.

With the Confluent MCP Server, an AI assistant works with Kafka topics, connectors, and Flink queries. This guide leads from the prerequisites to the first query.

Step 1: Check the prerequisites

You need Node.js 22.19.0 or later, an MCP client such as Claude Code or Cursor, and a Kafka environment. That can be a local cluster, a Confluent Platform, or a Confluent Cloud account.

Step 2: Generate the configuration

Run npx @confluentinc/mcp-confluent --init-config. This creates a config.yaml in the project folder. For a local cluster, add an entry under connections with bootstrap_servers and, if present, the Schema Registry endpoint. For Confluent Cloud, enter endpoints and API keys or choose browser sign-in with --init-oauth-config.

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

Step 3: Register it in the MCP client

In Claude Code claude mcp add confluent -- npx -y @confluentinc/mcp-confluent --config ./config.yaml is enough. In Cursor, add an entry to mcp.json with command npx and the same arguments. Do not put the config.yaml containing keys in a Git repository.

claude mcp add confluent -- npx -y @confluentinc/mcp-confluent --config ./config.yaml

Limit the tools

With --list-tools you see which tools are active. With --allow-tools you restrict the server to a selection, with --block-tools you block individual ones, such as delete-topics or produce-message. The allow list is applied before the block list.

Step 4: First queries

Start with read tasks, such as "List all topics" or "How large is the lag of the consumer group orders?". If a tool is missing, ask the assistant for explain-disabled-tools; it names the reason. Use write access only once the read tests look right.

Static API keys in config.yaml or OAuth (PKCE) via browser sign-in to Confluent Cloud for some of the tools; local clusters according to your Kafka configuration.

Read and write access to Kafka topics, messages, schemas, connectors, and Flink statements within the rights of the keys; can be limited with --allow-tools and --block-tools.

The server runs locally on your machine. Retrieved messages and metadata enter the AI client context and therefore reach the model provider. According to the README usage data and runtime errors (Sentry) are collected, which can be disabled with DO_NOT_TRACK=true.

Unintended deletion of topics, schemas, or connectors and sending of messages through write permissions, sensitive message contents in the model context, and API keys in an accidentally committed config.yaml.