Overview

kafka-mcp is the API + MCP server + A2A agent for Apache Kafka in the agent-utilities ecosystem. It wraps the Confluent REST Proxy (and, optionally, the native broker protocol) in typed, deterministic tools that an agent or policy router can call.

The tool surface

Six action-dispatch MCP tools cover the full Kafka surface. Each accepts an action plus a JSON params_json payload:

Tool	Resource	Representative actions
kafka_topics	Topics & configs	list, create, describe, delete, configs, update_config
kafka_partitions	Partitions	list, get
kafka_records	Produce / consume	produce, create_consumer, subscribe, consume, commit, delete_consumer
kafka_groups	Consumer groups	list, describe, consumers, lag_summary, lags
kafka_cluster	Clusters, brokers, ACLs	list_clusters, get_cluster, list_brokers, get_broker, broker_configs, list_acls, create_acl, delete_acls
kafka_native	Direct-to-broker	produce, consume, create_topic, delete_topic, list_topics, list_groups

The REST contract

KafkaApi targets the Confluent REST Proxy v3 for administration and v3 produce, with v2 helpers for the consumer-instance lifecycle:

• Clusters / topics / partitions / brokers / ACLs — GET|POST|DELETE v3/clusters/{cluster}/….
• Produce — POST v3/clusters/{cluster}/topics/{topic}/records with a typed value/key (STRING, JSON, BINARY, AVRO, JSONSCHEMA, PROTOBUF).
• Consume — the v2 instance lifecycle: create_consumer → subscribe → consume (poll) → commit → delete_consumer.

The active cluster id is resolved lazily and cached — honor KAFKA_CLUSTER_ID to pin it, otherwise the first cluster returned by list_clusters is used.

Where to go next

• Installation — pip, extras, Docker image.
• Deployment — transports, Compose, the agent server, Caddy, DNS.
• Usage — the tools, the KafkaApi client, the CLI.
• Architecture — the layered client and tool surface.