Jeston provides explicit transport and job primitives for applications that expose token output, progress, evaluations, or other long-running work. It does not include a model provider, train a model automatically, or guarantee model quality.
An API can return an AsyncIterable<Uint8Array> for NDJSON, SSE, token output, progress, or another byte protocol. React pages and API responses can also use React server streaming. This keeps the transport boundary deterministic while the application owns the workflow.
For agents, inference, evaluation, or training orchestration, jobs can submit or resume work. Applications must persist checkpoints and outputs in durable storage, use queue idempotency keys, propagate cancellation, and define failure recovery. A stream is not a durable job, and a durable job is not automatically a model evaluation.
AI workflows also require data governance, retention, consent, secret handling, evaluation, and provider-specific failure behavior. Jeston gives the application explicit HTTP, queue, cache, and deployment boundaries where those policies can be implemented and tested.
Jeston is useful here precisely because it does not hide the hard parts. It supplies deterministic boundaries; the application remains responsible for correctness. Read the source at github.com/kvantjs/jeston.
Top comments (0)