🦊 kafka-mcp

An MCP server that gives your AI agents eyes into Apache Kafka.

kafka-mcp exposes Kafka administration operations as Model Context Protocol tools, so assistants like Claude Desktop, Claude Code, or any MCP-compatible client can inspect and operate your cluster in plain language — "list all my topics with their replication factor" — instead of you reaching for the CLI.

It ships with a batteries-included docker-compose.yml that spins up a complete local Kafka lab (KRaft broker + Schema Registry + Web UI) so you can try it end-to-end in minutes.

✨ Features

• 🔌 Drop-in MCP server — runs over stdio, so any MCP client can launch it as a subprocess.

• 📋 Topic management — list & describe topics, create / delete, add partitions, and read or alter configs.

• 👥 Consumer group insight — list & describe groups, inspect members & assignments, and compute per-partition lag.

• 📨 Produce & peek — send a message to a topic, or read recent records back without committing offsets.

• 🩺 Cluster & offset views — describe brokers / controller and fetch earliest / latest watermarks per partition.

• ⚡ Async-friendly — blocking Kafka admin calls are offloaded to worker threads so the event loop stays snappy.

• 🐳 Self-contained local lab — one docker compose up gives you Kafka (KRaft, no ZooKeeper), Schema Registry, and a Web UI.

• 🛠️ Tiny & hackable — a single module (src/zaksway_kafka_mcp/__init__.py) you can read in one sitting and extend with new tools.

Related MCP server: Kafka MCP Server

🧭 How it works

┌──────────────────┐   MCP over stdio   ┌──────────────────┐   Kafka Admin API   ┌──────────────────┐
│   AI Agent       │ ◀───────────────▶  │   kafka-mcp      │ ◀────────────────▶  │   Kafka broker   │
│ (Claude, etc.)   │   tool calls       │  (FastMCP server)│   confluent-kafka   │  (localhost:9092)│
└──────────────────┘                    └──────────────────┘                     └──────────────────┘

The agent never talks to Kafka directly — it calls a tool, kafka-mcp translates that into a confluent-kafka admin or client request, and returns structured JSON the model can reason about.

📦 Prerequisites

• Python 3.12+

• uv for dependency management (recommended)

• Docker + Docker Compose (only if you want the local Kafka lab)

🚀 Quick start

1. Clone & install

git clone <your-repo-url> kafka-mcp
cd kafka-mcp
uv sync

2. Start a local Kafka (optional, but handy)

docker compose up -d

This brings up three services:

💡 Auto-create topics is enabled, so you can produce to a new topic and watch it appear via the MCP list_topics tool.

3. Run the MCP server

uv run kafka-zaksway

You should see:

Kafka MCP for you agents!

The server is now listening on stdio, ready for an MCP client to connect.

🤖 Connecting an MCP client

Most clients (Claude Desktop, Claude Code, …) launch MCP servers from a JSON config. Once it's installed from PyPI, point them at the published package — no clone required:

{
  "mcpServers": {
    "kafka-zaksway": {
      "command": "uvx",
      "args": ["zaksway-kafka-mcp"],
      "env": {
        "BOOTSTRAP_SERVER": "localhost:9092"
      }
    }
  }
}

• Claude Desktop → add the block to claude_desktop_config.json.

• Claude Code → claude mcp add kafka-zaksway -- uvx zaksway-kafka-mcp

💡 Hacking on a local clone instead? Swap to "command": "uv" with "args": ["--directory", "/absolute/path/to/zaksway-kafka-mcp", "run", "kafka-zaksway"].

Restart the client, and kafka-zaksway will appear among your available tools.

🧰 Available tools

kafka-mcp exposes 14 tools spanning topic management, consumer groups, the cluster, and the data plane. Tools marked ⚠️ are destructive (they delete data) — agents should confirm before calling them.

💡 The registered MCP tool names are full descriptive sentences (e.g. Show committed offsets and lag for a Kafka consumer group); the short identifiers above mirror the Python functions in src/zaksway_kafka_mcp/__init__.py and are used here for brevity.

Example — list_topics response:

[
  { "name": "orders",   "partitions": 6, "replication-factor": 1 },
  { "name": "payments", "partitions": 3, "replication-factor": 1 }
]

⚙️ Configuration

The server is configured entirely through environment variables.

📦 Releasing to PyPI

The package is published to PyPI as zaksway-kafka-mcp by a GitHub Actions workflow (.github/workflows/publish.yml) that triggers on v* version tags and authenticates via Trusted Publishing (OIDC) — no API tokens stored anywhere.

One-time setup — register a Trusted Publisher on PyPI:

To cut a release:

# 1. Bump `version` in pyproject.toml (e.g. 0.1.0 -> 0.2.0), then:
git commit -am "release: v0.2.0"
git tag v0.2.0
git push origin master --tags

The workflow verifies the tag matches pyproject.toml, builds the wheel + sdist, smoke-tests both, and publishes. Once published, anyone can run it with zero install:

uvx zaksway-kafka-mcp           # run the server directly
# or
pip install zaksway-kafka-mcp   # then run: kafka-zaksway

🗂️ Project structure

kafka-mcp/
├── src/zaksway_kafka_mcp/
│   ├── __init__.py     # The MCP server + all 14 tool definitions
│   └── __main__.py     # `python -m zaksway_kafka_mcp` entry point
├── tests/
│   └── smoke_test.py   # Import/packaging check run in CI before publish
├── .github/workflows/
│   └── publish.yml     # Build + publish to PyPI on `v*` tags (Trusted Publishing)
├── docker-compose.yml  # Local Kafka lab (broker + schema registry + UI)
├── pyproject.toml      # Project metadata, dependencies & build backend
├── uv.lock             # Pinned dependency lockfile
└── README.md           # You are here

🛣️ Roadmap

Recently shipped ✅

• create_topic / delete_topic

• add_partitions & alter_topic_config

• Describe consumer groups & their lag

• Peek at the latest messages on a topic

Ideas for what's next:

• Reset / set consumer group offsets

• ACL management (list / create / delete)

• Broker config inspection

• Schema Registry integration (list subjects & schemas)

🧑‍💻 Author

Zakaria BOUAZZA : https://zakaria.lu

📄 License

No license has been specified yet. Add one (e.g. MIT) before sharing publicly.