Camel 4.23 turns AI decisions and authorization into route primitives

Apache Camel 4.23 is more than a new long-term-support baseline. A detailed engineering walkthrough published after the release announcement shows the project moving AI decisions, workload identity and authorization into the same route model that integration teams already use for filters, choices and error handling.

Semantic decisions enter the route model

The release adds a preview semantic language whose evaluations can return a boolean, category, score or set of labels. Routes can reuse those results in Choice, Filter, Switch, Validate, Aggregate and error-handling steps, and can batch several evaluations into one call. Camel 4.23 ships adapters for TypeSafe AI’s decision API and for Wolf Defender, which detects prompt-injection and jailbreak attempts locally on a CPU through ONNX Runtime. Optional audit history is exposed through the CLI, developer console and Camel TUI.

That design matters because model-backed judgment is represented as a normal expression rather than buried in application glue. The important caveat is equally explicit: the semantic language is preview technology, and its YAML syntax changed during the 4.23 development cycle.

Authorization becomes part of the flow

Camel 4.23 also adds components for SPIFFE workload identity, Open Policy Agent and OpenFGA relationship checks. The release walkthrough says all three fail closed by default and can guard an entire route through security policies. AI tool routes gain an authorizationPolicy option, while MCP tools can return structured content and expose read-only resources.

For local tool integration, camel run ... --mcp-stdio can launch Camel as an MCP subprocess for an IDE or agent. Camel’s authoring MCP servers now share validation, catalog, file-editing and runtime tools, and repeated identical calls receive a short response intended to make an agent notice a loop.

A wider operational surface

The OpenAI-compatible component is also available through a vendor-neutral llm: scheme. New capabilities include moderation, image generation and editing, asynchronous batch operations, verified webhook consumption, token budgets and structured rate-limit metadata. A new observability module records LLM calls using OpenTelemetry GenAI semantic conventions and Micrometer metrics.

Outside AI, Camel 4.23 adds Kafka share-group consumption for Kafka 4.2 or newer. The component allows more consumers than partitions and acknowledges records independently, but the project warns that share groups do not provide ordering, offset seeking or transactions. The release also adds Switch and Cache EIPs, route conversion among YAML, XML and Java, and build-time validation for documentation examples.

What teams should evaluate

Integration teams should treat semantic routing as a preview and test its audit behavior, model dependency and failure modes before production use. Platform teams can evaluate SPIFFE, OPA and OpenFGA policies as reusable route controls rather than custom middleware. Users of vendor-packaged Camel distributions should wait for their supplier’s version and support guidance; the upstream 4.23 feature set does not by itself establish downstream availability.