Organization repository on GitHub
Vendor-neutral SDK for typed decision-making (yes/no, choice, score) — hosted and local
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Repos & SDKs
System One — asynq-io
Vendor-neutral Python SDK for System One decision models. One ask(state, questions) call over noul, choice and score, with provider choice expressed as a config class rather than a string. Hosted backends, a Fastino GLiDE preset, and an in-process ONNX path needing no torch at runtime.
- Stars
- 1
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- asynq-io
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is System One — iamaamir.
About this project
A vendor-neutral SDK for System One models. One contract — ask(state, questions) -> answers — over three question primitives (noul, choice, score). Switching providers is one config object, not a code change. These are decision models, not chat models: no messages, no streaming, no temperature. AsyncSystemOne has the same shape and the same method names, awaited: await agent.ask(...).
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- One contract, ask(state, questions) returning answers, plus AsyncSystemOne with the same method names simply awaited.
- Provider choice is a config class, not a string: TypesafeConfig, OpenRouterConfig, HTTPConfig or ONNXConfig, each overridable by SYSTEM_ONE_* variables.
- The ONNX backend runs in process with no key and no network; system-one fetch writes the graph, a companion config and the tokenizer directory.
- Retries default to two and cover 408, 429 and 5xx only; 401 and 403 raise AuthenticationError and are never retried.
Watch out
The abstraction is deliberately narrow: no messages, no streaming and no temperature, which suits decisions and does nothing for generation. The local path is only as good as laya, since system-one fetch pulls convaiinnovations/laya and fp32 is around 1.6 GB; any other variant or your own model needs a YAML spec and the export extra. A default pytest run does not exercise the local backend, because tests needing onnx/laya.onnx skip themselves when it is absent and parity against the reference laya implementation is a separate script requiring the export extra and the real weights. Retry policy ignores everything outside 408, 429 and 5xx, and a malformed body surfaces as pydantic's ValidationError rather than an SDK error class. Fastino's idle models cold-start, which is why its example sets a 300-second timeout. 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.
A config class rather than a string, over a deliberately thin contract: ask(state, questions) with noul, choice and score, and no messages, streaming or temperature. Anything vendor-specific stays reachable on agent.backend rather than being modeled, and tests can inject a fake backend directly through SystemOne(using=FakeBackend()).
Yes through the ONNX backend, but the weights are someone else's. system-one fetch --out-dir onnx pulls convaiinnovations/laya and writes onnx/laya.onnx at roughly 1.6 GB fp32 plus a config and tokenizer, then runs in process with no network, no key and no torch at runtime. A different laya variant or your own model means writing a spec and running system-one export, which needs the export extra.
No. Tests needing the ONNX artifacts skip themselves when onnx/laya.onnx is missing, with SYSTEM_ONE_ONNX_DIR and SYSTEM_ONE_MODEL overriding where they look. Checking the graph against the reference laya implementation is a separate script, scripts/check_onnx_parity.py, which needs the export extra and the real weights, so a green default run says little about the local backend.
Retries default to two and apply to 408, 429 and 5xx only, governed by SYSTEM_ONE_MAX_RETRIES and SYSTEM_ONE_TIMEOUT, which defaults to 10 seconds. A 401 or 403 raises AuthenticationError and is never retried. APIError carries status, body and retry_after; a malformed response body comes back as pydantic's ValidationError rather than one of the SDK's own errors.
As extras. pip install "system-one[http]" pulls the hosted-provider side and pip install "system-one[onnx]" the local in-process backend; system-one fetch additionally wants the hub extra. Exporting your own model is a third extra again, so a local-only install that does not build graphs can stop at the onnx extra.
Answers split by primitive: response.nouls["urgent"].noul with its confidence, response.choices["topic"].choice alongside the full probabilities, and response.scores["severity"].score. AsyncSystemOne has exactly the same shape and the same method names, just awaited, so sync and async code read the same.
TypesafeConfig for hosted Jev reading SYSTEM_ONE_API_KEY, OpenRouterConfig for the same body against a different endpoint, and plain HTTPConfig for any host plus model you name. Fastino's GLiDE is configured through HTTPConfig at https://api.fastino.ai with model fastino/GLiDE, where the key is accepted as a bearer token or an X-API-Key header, and the shipped example sets a 300-second timeout because idle models cold-start.
More Jev resources like this
45GitHub stars
Go SDK plus a pipeline-friendly CLI for typed decisions, covering single questions, batched question maps, retries, and line-oriented filtering over stdin.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
mattn/go-jevGo
mattn/go-jevGo5GitHub stars
Zero-dependency Node bridge that exposes Jev through OpenAI-style chat and responses endpoints with streaming, plus a CLI, a code-audit tool, and an auto-fix loop.
Jev Repos & SDKs#community#awesomejev#catalog-import
6GitHub stars
A synchronizer that watches the official TypeSafe docs, agent skill and SDK repositories and regenerates a provenance-tracked reference: current state, semantic events and human-readable change reports.
Jev Repos & SDKs#community#awesomejev#catalog-import
Community
chenrui333/jev-docsPython
chenrui333/jev-docsPython1GitHub stars
A Python layer over a small decision-provider contract that adds validated distributions, batching, budgets, ranking and exportable records, with Jev as the first provider.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
yairshy/decidoPython
yairshy/decidoPython1GitHub stars
Community Go client for the TypeSafe API: context-aware SystemOne calls, Choice, Noul and Score constructors, scoped functional options, typed errors carrying RFC 6901 field paths, and standard-library-only runtime code. Requires Go 1.24.
Jev Repos & SDKs#community#awesomejev#catalog-import
Community
unimtx/typesafe-sdk-goGo
unimtx/typesafe-sdk-goGo1GitHub stars
C# client for Jev, TypeSafe AI's System One model. Typed Noul, Choice and Score questions go out with one state in a single round trip and come back with a value, a probability distribution and a confidence number. Choice questions can be built from an enum, so the compiler finds the switch branch you forgot.
Jev Repos & SDKs#community#awesomejev#catalog-import
Community
RavenValentin/TypeSafe.JevC#
RavenValentin/TypeSafe.JevC#From the Jev guides
Original write-ups on what this entry is for.
Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.
Jev Repos & SDKs#open-models#research#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.