<- dungtd / blog_
VI
$ cat ./blog/gen-ai/he-sinh-thai-jev.md
#decision-model#llm#production
seriesDecision Models: When AI Decides Instead of Writespart 3

The JEV Ecosystem: Official SDKs, Community Projects and Open System One Models

Part 3 of the Decision Models series: a list of the tools and repositories around JEV, grouped into three layers — official (the typesafe-ai org), community, and open / self-hosted. Each tool gets a name, a link and a one-line description. Repositories checked on 30 September 2026.

date2026-09-30read5 minwords1,000byTrương Đức Dũng
~/diagrams/he-sinh-thai-jev/jev-ecosystem-map-en.htmlinteractive
open ↗
FigThe official path runs from JEV (a closed model with no published weights) through the TypeSafe API at POST /v1/systemone to the official SDKs (Python, JS and an LLM-backed adapter), then to official integrations such as the n8n node and skills. Community projects wrap the SDK in their own integrations (Spring AI, neurolink) and applications (browser agent, guard, search). The open-model branch runs open models such as Qwen, ModernBERT and Gemma behind a compatible server (Kev, LocalJev, ollaya) that speaks the same protocol, so the official SDK can point at it. The figure only groups projects; the full lists are in sections 01–03.
official-layer node (TypeSafe)community-layer nodeopen / self-hosted (grey box, dashed purple edge)

In part 2, we opened up a single JEV call from the inside: the request, the typed response with probabilities, and the policy code. This post looks outward: which tools and repositories are being built around JEV.

The list is split along one axis: who is responsible for the repository. When you pick something to use, the first question is who fixes it when it breaks, so the three layers below call for three different levels of trust:

  1. Official: repositories in the typesafe-ai org, maintained by TypeSafe.
  2. Community: projects outside TypeSafe that call the real JEV through the SDK or HTTP.
  3. Open / self-hosted: not calling JEV, but running an open model behind a server that speaks the same POST /v1/systemone protocol.

Figure 1 at the top of the post draws these three layers. Each description follows the repository’s own README or description, checked on 30 September 2026; the blog has not verified the claims in them.

01The official layer

Repositories in the typesafe-ai org that TypeSafe’s official docs and PyPI page point to:

02The community layer

TypeSafe Community (TypeSafeAI)

The TypeSafeAI org has a name close to the official one but describes itself as an “UNOFFICIAL organization made for community repos”, so it belongs in the community layer.

  • typesafe-router — picks a model or tool with a single Choice question.
  • jev-harness — a harness for coding agents: the LLM proposes, JEV answers narrow questions about the proposal, code decides.
  • typesafe-playground — a playground with many use cases.
  • clarity-judge — grades writing quality along several axes.
  • typesafe-ui — a UI component kit.

Applications

SDKs, integrations and CLIs

Curated lists

03Open and self-hosted models

This branch runs open models in place of JEV. The servers below speak the same protocol, so the official SDK can point at them through base_url, but the model behind them is not JEV.

Compatible servers and runtimes

  • jaredpalmer/kev — self-trained decision models on Qwen (0.8B–27B), with weights and a compatible server.
  • githubnext/localjev — a TypeScript/Bun bridge from GitHub Next that returns JEV-shaped responses from DiffusionGemma.
  • razorback16/openjev — a JEV-compatible server on DiffusionGemma.
  • ekzhang/openjev-sglang — a JEV-compatible endpoint on open models, prefill only.
  • ollaya-dev/ollaya — runs open decision models locally behind a TypeSafe-compatible API, the way Ollama runs LLMs.

Open models

04Summary

  • Official is just six repositories in the typesafe-ai org: two SDKs, the LLM adapter, skills, the n8n node and WorkflowEvals. Check whether something is official by going from TypeSafe’s docs and PyPI out to the repository.
  • Community covers the TypeSafe Community org (unofficial) and other groups, calling the real JEV to pick actions, score context, score risk or choose sources.
  • Open / self-hosted covers compatible servers and open models; same protocol but not JEV, so thresholds chosen on JEV must be measured again when you switch.

The next post, part 4, goes from this list to a concrete application: designing a Chrome extension for web data extraction with JEV, in the same way jev-ultrafast lets JEV only pick while code executes. If you have not read the earlier posts, part 1 explains why this class of model exists, and part 2 covers the schema, the three question types and how to turn probabilities into actions.

References

Official TypeSafe AI sources

Community and open-model repositories

  • Links and descriptions for each repository in sections 01–03 follow that repository’s own README, checked 30 September 2026.
  • TypeSafe Community — TypeSafeAI org (“UNOFFICIAL organization” description)

Author’s notes

  • Original notes: “JEV (TypeSafe AI) — Toàn cảnh ứng dụng & mã nguồn mở GitHub” (Notion). Corrections made while writing: the TypeSafeAI org moved to the community layer, n8n-nodes-typesafe-ai and WorkflowEvals added to the official layer. Apps that have a name but no link are left out.

Trương Đức Dũng

AI Engineer - Hanoi

Focused on LLMs, RAG, and AI systems that need to run reliably in production.

profile ->github ↗