# Confluent MCP — Kafka + Flink Tools via MCP

> Confluent MCP Server exposes 50+ Confluent tools (Kafka, Flink SQL, Schema Registry and more) to MCP clients, with a config-driven setup via npx.

## Install

Merge the JSON below into your `.mcp.json`:

## Quick Use

1. Generate a starter config:
   ```bash
   npx @confluentinc/mcp-confluent --init-config
   ```
2. Edit `config.yaml`, then run:
   ```bash
   npx @confluentinc/mcp-confluent --config ./config.yaml
   ```
3. Optional: list tools:
   ```bash
   npx -y @confluentinc/mcp-confluent --list-tools
   ```

## Intro

Confluent MCP Server exposes 50+ Confluent tools (Kafka, Flink SQL, Schema Registry and more) to MCP clients, with a config-driven setup via npx.

- **Best for:** teams who want agents to inspect/operate Kafka + Confluent Cloud with a controlled allow/block list
- **Works with:** Node.js (per README); Confluent Cloud/Local; MCP clients (Claude Desktop/Code, Cursor, VS Code, Gemini CLI)
- **Setup time:** 15–30 minutes

## Practical Notes

- Quant: the README describes **50+ tools** across Kafka, Flink SQL, Schema Registry and more.
- Quant: you can validate tool surface quickly with `--list-tools` before wiring it into an MCP client.

## A safe integration pattern

The fastest way to get burned is to give an agent a giant, write-capable surface area on day 1. Instead:

1. **Start with discovery-only tools**: list clusters/topics, read configs, fetch schemas.
2. **Add a narrow allow-list** for write actions (produce, create topics) only after you have a human review loop.
3. **Version and review config.yaml** like code: PRs, diffs, and environment scoping.

## What to measure

- Time to answer “where is this message coming from?” before/after MCP adoption.
- Number of manual CLI steps replaced by agent tool calls (start with 5 per incident).

### FAQ

**Q: Do I need to install globally?**
A: No. The README shows running via `npx` with a config file.

**Q: How do I restrict actions?**
A: Use allow/block lists (see CLI options) and start with read-only exploration.

**Q: What is a good first task?**
A: List environments/clusters/topics, then inspect one schema and consumer group flow end-to-end.

## Source & Thanks

> Source: https://github.com/confluentinc/mcp-confluent
> License: MIT
> GitHub stars: 152 · forks: 50

---

<!-- ZH -->

## 快速使用

1. 生成初始配置：
   ```bash
   npx @confluentinc/mcp-confluent --init-config
   ```
2. 编辑 `config.yaml` 后启动：
   ```bash
   npx @confluentinc/mcp-confluent --config ./config.yaml
   ```
3. 可选：列出可用工具：
   ```bash
   npx -y @confluentinc/mcp-confluent --list-tools
   ```

## 简介

Confluent MCP Server 将 Kafka、Flink SQL、Schema Registry 等 50+ Confluent 工具暴露给 MCP 客户端；用 npx 生成并加载 config.yaml 即可启动。

- **适合谁：** 希望让 Agent 以“可控白/黑名单”的方式读写 Kafka/Confluent Cloud 的团队
- **可搭配：** Node.js（见 README）；Confluent Cloud/Local；MCP 客户端（Claude Desktop/Code、Cursor、VS Code、Gemini CLI 等）
- **准备时间：** 15–30 分钟

## 实战建议

- 量化信息：README 描述了覆盖 Kafka、Flink SQL、Schema Registry 等的 **50+ tools**。
- 量化信息：可先用 `--list-tools` 校验工具面，再接入 MCP 客户端，减少踩坑成本。

## 更安全的接入姿势

最容易翻车的做法：第一天就把“可写的巨大工具面”交给 Agent。更稳妥的方式：

1. **先只读探索**：列出集群/Topic、读取配置、拉取 Schema。
2. **写操作走窄白名单**：produce / create topic 等动作先接入人工评审流程再逐步放开。
3. **把 config.yaml 当代码管理**：走 PR、做 diff、按环境拆分，避免一份配置通吃所有环境。

## 建议量化指标

- 接入前后，“定位消息来源/异常 Topic”的耗时变化。
- 每次故障中，手工 CLI 步骤被 Agent tool call 替代的数量（先以每次 5 步为目标）。

### FAQ

**一定要全局安装吗？**
答：不需要。README 的示例就是用 `npx` 搭配配置文件运行。

**怎么限制危险操作？**
答：用 allow/block 列表，并从只读探索开始逐步放开。

**第一件事做什么？**
答：先列出环境/集群/Topic，再选一个业务链路把 Schema→生产→消费走通。

## 来源与感谢

> Source: https://github.com/confluentinc/mcp-confluent
> License: MIT
> GitHub stars: 152 · forks: 50


---
Source: https://tokrepo.com/en/workflows/confluent-mcp-kafka-flink-tools-via-mcp
Author: MCP Hub