Production notes & limitations
What v0.1.0 is, what it is not yet, and how to think about running it.
v0.1.0 is an early, open-source release. It is honest about what it does and does not do yet.
What ships in v0.1.0
- The single-system runtime: journal, content store, commit protocol, refusal gate, CapabilityBroker.
- The
kxCLI (FFI-free single binary) andkx serve(embedded gRPC gateway + event WebSocket). - The five recipes.
- Optional local inference (Tier 1) and the optional coordinator/worker layer.
Limitations to plan around
- Early release. v0.1.0 is intentionally scoped. It is not "battle-tested" or "production-hardened" — treat it as a capable foundation, not a turnkey SLA.
- Single-node by default. The distributed layer is optional and you opt into it deliberately.
- Local inference needs a C++ toolchain and is not part of the FFI-free default build. CUDA is cloud-only.
- Auth is deny-all by default on
kx serve— you must explicitly enable--dev-allow-localor pass an--auth-token.
Operating posture
- Treat the journal as the source of truth — back it up; everything else (projections) is re-derivable.
- Lean on the refusal gate and CapabilityBroker for policy: effects only happen through the single door.
- Use the exactly-once proof as a recovery smoke test in your own environment.
Honest by design
We don't claim production-ready. v0.1.0 ships now so you can build on it in the open — the platform features (hosted, multi-agent orchestration, dashboard) are the next phase.
Extending
Seven trait seams make the runtime swappable — Journal, ContentStore, CapabilityBroker, InferenceBackend, ResourceManager, SecretStore, WorkerRegistry.
Roadmap
What's shipping after v0.1.0 — client SDKs, a dashboard, metrics export, multi-modal inference, a model-driven loop, and an optional managed cluster layer.