Is this an official TypeSafe SDK, and what happens when it drifts from the API?

It is a community client and the README states that in the first paragraph. The mitigation is that the decoded JSON body is kept in the raw field, so an answer type the installed version does not know about is skipped with a warning instead of crashing, and stays readable there. You would be reading raw maps for anything new upstream adds until a release handles it.

How do I test code that calls Jev without spending tokens?

Point the client at Req.Test with req_options: [plug: {Req.Test, TypeSafe}] in config/test.exs, then stub per-test answers with TypeSafe.Test.stub_answers/1, for example {:choice, :technical, %{billing: 0.08, technical: 0.85, sales: 0.07}, 0.82}. Set retry: false in the same config so a stubbed failure does not retry three times. Tests run async and never touch a real key.

Will a 429 turn into a retry storm against my rate limit?

The default policy makes two retries after the first attempt with exponential backoff from 500 ms to a 5 s cap and 25% jitter, retries 408, 429 and every 5xx including 529, plus connection failures and timeouts, honours Retry-After, and gives the whole call a 30 s budget. You can widen it with a keyword list such as retry: [max_retries: 4, budget_ms: 60_000], or switch it off per call with retry: false.

Where does my API key actually come from, and can I override the model per call?

Every option resolves in a fixed order: the value passed to TypeSafe.new/1, then the matching TYPESAFE_* environment variable with blanks ignored, then config :typesafe, then the default. The default base URL is https://api.typesafe.ai and the default model is the jev-latest alias, which resolves to a concrete version that is reported back in the response. ask/4 also takes per-call :model, :timeout, :retry, :headers and :telemetry_metadata.

What are the Elixir and OTP requirements?

Elixir 1.18 or later for the standard library JSON module, and Erlang/OTP 27 or later, with Req, Zoi, and Telemetry as runtime dependencies.

How are Score answers keyed?

Score legend and probabilities are keyed by integer level, so you can write probabilities[2]; Choice and Noul answers come back under your ids and keep your key type, with atoms staying atoms.

What telemetry does each call emit?

One [:typesafe, :request, :start] event and one :stop event, or :exception; the :stop event carries duration, token usage, status, request id, retry count, and any error.

CommunityJev Repos & SDKs

TypeSafe — mattneel

An unofficial Elixir client for TypeSafe's System One API. Noul, Choice and Score questions go out in one request and come back as structs with probabilities, and thresholds and policy stay in your code.

See highlights

Open github.com

Stars: 2
Language: Elixir
License: MIT
Published by: Community
Author: mattneel
Last verified: Oct 2026

mattneel/typesafe

User repository on GitHub

View on GitHub

An idiomatic Elixir client for the TypeSafe AI API

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is TypeSafe — krzyzanowskim.

About this project

Elixir client for TypeSafe's System One API and its Jev model. Ask typed Choice, Noul and Score questions about your application's state in one request, and get back structs with probabilities that your code acts on. This is a community client, not an official TypeSafe SDK. The SDK returns judgments as data. Thresholds and policy stay in your code. Add typesafe to your dependencies in mix.exs:

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: Elixir

Highlights

Answers come back keyed by your own ids: atom ids and atom Choice options stay atoms, strings stay strings.
A %TypeSafe.Error{} struct covers every failure mode, and invalid_request errors carry a :path pointing at the offending value.
The default retry policy matches TypeSafe's Python and JavaScript SDKs: two retries, exponential backoff, Retry-After honoured, 30 s budget.
Tests route through Req.Test with stub_answers and run async, so no test needs a real key or a paid call.

Watch out

This is a community client and the README says so directly — it is not an official TypeSafe SDK, so it can lag the API. It pins Elixir 1.18 or later (for the standard library JSON module) and Erlang/OTP 27 or later, which is a higher floor than the Node SDK needs, and it sits at version 0.1.0. It depends on Req at ~> 0.7.4 or ~> 0.8 plus Zoi and Telemetry, and the test helpers need Plug on top of that. The client is an immutable value with no process state, so you build it once and share it; the Finch pool owns the connect timeout, not the :timeout option, which only covers receiving. :retry_delay is rejected while :retry is a policy struct, and TYPESAFE_LOG_LEVEL is not read — you configure Logger yourself. An answer type this SDK version does not recognise is skipped with a warning and stays reachable in raw, so a schema addition upstream becomes a silent gap rather than an error. 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_ex

An unofficial Elixir SDK for TypeSafe's System One endpoint, built on Req and Finch, with per-call retry policies, a shared connection pool and status-mapped errors.