Is this official TypeSafe code?

No. It is an unofficial client maintained by Kunobi, and the README states it is not affiliated with TypeSafe. The value it adds is portability and ergonomics: one typed request and result shape across several providers, plus enums, confidence helpers and a test double, rather than anything TypeSafe does not already offer.

Does batching many questions into one call actually save money?

Yes, according to a measurement the README reports against the live API: a 24k-token state with one question cost 24,288 input tokens, and the same state with 200 questions cost 26,676, while latency stayed flat. The reason is that the state is charged once however many questions ride along, whereas one question per call re-sends the state every time.

How do I test code that calls this without hitting the API?

Depend on the trait rather than the concrete client, taking &dyn SystemOne or an impl SystemOne instead of Client, and enable the testing feature. FakeSystemOne answers from a script, checks each scripted answer against the question it was asked, and records the requests it received, so you can assert on what your code would have sent.

How do I avoid auto-routing an uncertain ticket to the wrong team?

Use the threshold helpers rather than reading raw probabilities. decide(min_confidence) returns the label only when confidence reaches your threshold, and scores and yes/no answers have decide and is_yes too, so you can send low-confidence cases to a person. A label the enum does not know about fails loudly with Error::UnexpectedAnswer instead of silently matching.

Which providers can this Rust client talk to?

TypeSafe by default, plus Cloudflare Workers AI (clef), Liquid d1, Vercel AI Gateway, OpenRouter, and any compatible System One endpoint, each with its own key environment variable.

What Rust version and features does the client need?

MSRV 1.94; calls run on Tokio, with optional native-tls, a blocking client for callers with no async runtime, and a testing feature that adds a scripted FakeSystemOne.

What limits does the client put on questions?

A Choice takes up to 255 options and a Score between 2 and 10 levels, and the client refuses more before sending; ScoreAnswer::normalized() and ranked() help combine and order them.

CommunityJev Repos & SDKs

Kunobi Jev

A Rust client for typed decision models across TypeSafe Jev, Cloudflare Clef, Liquid, Vercel AI Gateway and OpenRouter, with noul, choice and score builders, typed label enums, and a scripted fake for tests.

See highlights

Open github.com

Stars: 1
Language: Rust
License: APACHE-2.0
Published by: Community
Author: kunobi-ninja
Last verified: Oct 2026

kunobi-ninja/kunobi-jev

Organization repository on GitHub

View on GitHub

Rust client for the TypeSafe System One API (Jev)

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Rust client for typed decision models: TypeSafe Jev, Cloudflare Clef, Liquid d1, Vercel AI Gateway and OpenRouter. Send state and typed questions, and get answers your code can act on: a label, a score or a probability, each with the numbers behind it. The client supports three question types:

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Rust

Highlights

Questions are built with noul, choice and score helpers; choice labels can be declared as a Rust enum and matched on instead of on strings.
Batching is the cheap path: a 24k-token state cost 24,288 input tokens with one question and 26,676 with 200, latency flat.
decide(min_confidence) returns a label only when confidence reaches the threshold, so uncertain cases can be routed to a person instead of guessed.
Take &dyn SystemOne instead of Client and use FakeSystemOne: scripted answers are checked and requests recorded.

Watch out

This is an unofficial client maintained by its author and explicitly not affiliated with TypeSafe, so treat its provider behaviour as third-party. The batching numbers are one measurement against the live API rather than a published benchmark, and provider support is uneven in ways the README is upfront about: Liquid's documentation specifies no model catalog, so listing models there costs input tokens on a probe request, and OpenRouter uses no vendor or model-name allowlist, meaning anything advertising the decisions output is selectable. Only TypeSafe reads its base-URL and model environment variables, which is deliberate isolation but surprises anyone expecting environment config to carry across providers. The default HTTP client refuses redirects and the base URL must be https. Note the crate was renamed from kunobi-jev to kunobi-decision, with a compat crate re-exporting the old types for staged migration. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

typesafe-sdk-rust

Unofficial Rust client for the TypeSafe AI System One API, matching the Python SDK 0.6.0 contract: named noul, choice and score questions, a typestated builder, per-call overrides, and a `wire` module for serving the same shape.