What is not built yet

Product guide: what the capability brings, for whom, its limits and the version that delivered it - no configuration syntax.

Ordered by planned version.

Milestone	What comes
Performance measurements	Benchmarks of the native Kafka proxy and of SLO objectives on groups (released in 1.6.0): written, not run yet - no figure is promised before them
No announced version	Native Kafka clients: reads counted in the tenant's daily or monthly budget; SLO objectives on groups: end-to-end proof under an injected outage
1.7.0 IoT and industry	MQTT devices connecting to the gateway (a proxy in front of your broker), device identity, Sparkplug B, OPC UA PubSub JSON, a devices console
1.8.0 enterprise needs	Verified incoming webhooks (HMAC, Standard Webhooks, Stripe, GitHub, Slack, Shopify, SendGrid), IP address lists, plans, AI budgets per user and per use case, negotiated compression, read coalescing, a plans console; traffic split between versions of a service (canary, blue-green, rollback); AI costs per application, team, project and cost centre, tokens per minute, thresholds (warn, switch to a cheaper model, refuse), budget alerts, forecast, chargeback export, OpenTelemetry conventions for generative AI
1.9.0 governed AI	Routing by declared task (a small model for a simple task, a reasoning model for an analysis), choice by hosting, region, cost or measured latency, model canary, allowed models per team or plan; data protection: data classes and models allowed per class, residency, detection and masking of personal data in prompts, injection patterns, decisions traced without the content; inference-aware routing, semantic cache, signed A2A cards and registry, execution proof (EU AI Act)
1.10.0	Aggregation of several upstreams, automatic proposal rules, AI for operations; agent actions: verified human approval for a sensitive tool, rights limited to one task, argument checks by identity, a budget per agent run; an AI services catalogue
1.11.0	Developer portal (catalogue, try-out, keys - the developer guide prefigures it), traced emergency break-glass, optional Kubernetes operator; edge mode: the same gateway on a site (plant, farm, facility), pulling a signed configuration and deciding without a network, buffering left to your local broker
No milestone	Multi-gateway console; sender-constrained tokens (DPoP, mTLS RFC 8705); Sunset and Deprecation headers toward callers; licence (to be dated); candidates on real demand: shared policies across gateways, persisted GraphQL queries, CoAP, commands gated by a system's state, FHIR profiles, cross-organisation exchanges, AI providers without an OpenAI-compatible API
2.0.0 (on real demand)	WebAssembly policies, HTTP/3 on the caller side, XML, YAML and form decoders

Not planned: public TLS termination, WAF, anti-DDoS; issuing tokens for callers; a model that decides; a native plugin in the process; a console that changes the gateway; a configuration language beside the contract; terminating MQTT or serving static files; an agent runtime, a RAG platform, response quality evaluation, a prompt registry, an approval workflow engine; Modbus, OPC UA client-server and buffering a site's messages (shop-floor gateway and local broker).