A Stateless Rosetta Stone for Kafka and AI Agents

What it does Zilla is a stateless, multi-protocol gateway built on a streaming-native engine. It exposes Apache Kafka and MQTT to applications over HTTP, SSE, gRPC, MQTT, or WebSocket, and it federates MCP servers, HTTP APIs, OpenAPI services, and Kafka behind a single agent-facing endpoint. Everything is declared in one zilla.yaml file, sharing routing, identity, authorization, and telemetry across both event and AI traffic.

The interesting bit The project treats protocol translation as a low-level streaming problem rather than a layer of adapters: it uses code-generated flyweights over encoded buffers, pins each connection to a single worker, and passes back-pressured frames through shared memory. The newer MCP Gateway angle is particularly unusual—it namespaces capabilities like github__create_pr or kafka__produce_message so AI agents see one catalog while Zilla routes each call to the correct backend, handling auth and credential exchange itself.

Key highlights

• Exposes Kafka to protocol-agnostic clients (browsers, IoT devices) without custom bridges.
• Federates multiple MCP providers, HTTP APIs, and OpenAPI services behind one governed endpoint.
• Namespaces capabilities with a toolkit__capability scheme to avoid collisions and simplify routing.
• Supports eager and cold tool discovery so AI agents load only the capabilities they need.
• Stateless request processing with shared state externalized for multi-replica deployments.

Caveats

• Advanced OAuth grants and distributed stores (Redis/Hazelcast) are gated behind the commercial Zilla Plus edition.
• The Aklivity Community License allows production use but explicitly prohibits offering Zilla itself as a standalone commercial service.

Verdict Worth evaluating if you are tired of maintaining bespoke Kafka-to-HTTP glue or need to govern AI agent access to disparate APIs and event streams. Skip it if you already have a simple, single-protocol proxy and no MCP or multi-protocol translation needs.