NotJev : Kev

Typed decisions for software.
Send context. Ask choice, score, or noul.
Get calibrated probabilities — not a paragraph to parse.

Install · Docs site · Quick start · Benchmarks · Why Kev · Playground · SDKs · Migrate from Jev

Install

npm install @kev-ai/sdk
npm install -g @kev-ai/server @kev-ai/cli

Package	What	npm
`@kev-ai/sdk`	TypeScript client (`choice` / `score` / `noul`)	npmjs.com/package/@kev-ai/sdk
`@kev-ai/server`	Decision API + playground → `kev-server`	npmjs.com/package/@kev-ai/server
`@kev-ai/cli`	CLI + MCP → `kev` / `kev-mcp`	npmjs.com/package/@kev-ai/cli

kev-server    # http://127.0.0.1:3000  (+ /playground/)
kev demo
kev health

LangChain / LlamaIndex: @kev-ai/sdk/langchain · @kev-ai/sdk/llamaindex
PyPI coming later.

curl -s http://127.0.0.1:3000/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{
    "state": "Charged twice. I am furious.",
    "questions": {
      "topic": {
        "type": "choice",
        "instructions": "Which team?",
        "criteria": {
          "billing": "charges and refunds",
          "technical": "bugs and outages"
        }
      },
      "escalate": {
        "type": "noul",
        "instructions": "Escalate to a human now?"
      }
    }
  }'

{
  "answers": {
    "topic": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.97,
      "probabilities": { "billing": 0.97, "technical": 0.03 }
    },
    "escalate": { "type": "noul", "noul": 0.86 }
  }
}

if (answers.escalate.noul > 0.7) routeToHuman();
else assign(answers.topic.choice);

NotJev : Kev is an open-source System One decision engine: Apache-2.0, self-hosted, Jev-style wire format. Bring your own model (Ollama, vLLM, OpenAI-compatible) — or start with the built-in mock.

Independent project. Not affiliated with TypeSafe AI or OpenJev.

Docs & website

Product docs, benchmark write-ups, and a fixed Try it workbench live in the companion Next.js site (sibling repo / folder NotJev-Kev-Website):

cd ../NotJev-Kev-Website   # or clone your website repo
npm install && npm run dev

Routes: / · /docs · /docs/benchmarks · /docs/why · /try (showcase workbench).
GitHub on that site always points here — engine source stays in this repo.

What NotJev : Kev is

NotJev : Kev turns an LLM into a typed classifier / gate / scorer your app can trust:

Primitive	Returns	Use for
`choice`	Winning key + full distribution	Routing, intent, triage (≤255 options; cascade for larger)
`score`	Ordinal level + distribution	Urgency, quality, sentiment
`noul`	Probability in `[0, 1]`	Escalate? Toxic? Block? Page on-call?

You do not train models. You do not parse chat. You threshold numbers.

Why NotJev : Kev

vs chat LLMs

You need	Chat gives you	Kev gives you
A label	Prose you must regex	A typed key + probabilities
A gate	“I think so…”	`noul` you can threshold
Confidence	Vibes	Concentration of the distribution
Many judgments	N serial prompts	One request, questions in parallel
Control	Vendor lock-in	Your GPU / API / laptop

What makes NotJev : Kev different

	Hosted Jev / OpenJev	NotJev : Kev
License	Proprietary / mixed	Apache-2.0
Deploy	Their cloud / their weights	Self-host anywhere
Models	Fixed stack	BYO — Ollama, vLLM, OpenAI, Gemini-compat, …
Wire format	System One	Same shape (`choice` / `score` / `noul`)
Offline	No	Mock backend for CI & demos
DX	API key	Playground · SDK · CLI · MCP · LangChain / LlamaIndex
Training	N/A for you	None — inference + API only

NotJev : Kev is the open control plane. The intelligence is whatever model you point it at.

Benchmarks

Same public held-out suite used around OpenJev: s1lv3rj1nx/openjev-heldout.
Model: qwen3.5:9b via Ollama · 100% GPU.

Task	Primitive	Kev	Chance	Reference
Banking77	choice (77) · n=600	83%	1.3%	Jev ~82% held-out · JevBench 80.3%
CLINC OOS	choice (151) · n=600	86%	0.7%	—
AG News	choice (4) · n=600	86.5%	25.0%	—
SST-5	score (5) · n=600	89.5% within-1	20.0%	ordinal ±1
Civil Comments	noul · n=600	81%	50.0%	—

OpenJev’s private 10k mix (cite only): Jev 85.4% · OpenJev 84.0% — questions not fully public, so we don’t invent a fake score against it.

Ops metrics (speed · structure · stability · batching)

Kev’s preferred path is letter-token readout when the backend exposes logprobs; on Ollama that falls back to constrained decode (still first-shot, no JSON retries). Measured on qwen3.5:9b · 100% GPU:

Metric	`qwen3.5:9b` (warm GPU)
p50 / p95 (2 questions)	504 / 803 ms
Parse-fail / format hallucination	0%
Rerun agreement flip / KL	0% / 0
Multi-Q @15 parse-fail	0%
Multi-Q @1 → @15 p50	443 → 8011 ms (~linear)
Confidence @15 vs @1	~1.17× (no drop)

# Accuracy (held-out)
pnpm bench:heldout -- --mode api --base-url http://127.0.0.1:3000 \
  --tasks banking77,clinc_oos,ag_news,sst5,civil_comments_toxicity

# Speed + agreement + multi-Q
pnpm bench:ops -- --mode api --base-url http://127.0.0.1:3000 --trials 10

Details: benchmarks/RESULTS.md · benchmarks/METHODOLOGY.md

Quick start

Requirements: Node 20+, pnpm 9.

git clone https://github.com/arjun988/Kev.git
cd Kev
pnpm install
pnpm build
cp .env.example .env
pnpm --filter @kev-ai/server start

Open	URL
Playground	http://127.0.0.1:3000/playground/
Health	http://127.0.0.1:3000/health
OpenAPI	http://127.0.0.1:3000/openapi.json

Use a real local model (Ollama)

ollama pull qwen3.5:9b   # or llama3.2, etc.

KEV_BACKEND=ollama
KEV_OLLAMA_MODEL=qwen3.5:9b

Restart the server. Confirm GPU with ollama ps → 100% GPU.

Windows: corepack enable fails with EPERM?

iwr https://get.pnpm.io/install.ps1 -useb | iex
# restart the terminal, then pnpm install && pnpm build

Python SDK

cd python && pip install -e . && cd ..

Playground

A built-in UI for support routing, moderation, content scoring, vision triage, and intent — with probability bars and raw JSON.

→ http://127.0.0.1:3000/playground/

Use it in code

Published packages (npm org @kev-ai · PyPI kev):

npm install @kev-ai/sdk
npm install -g @kev-ai/server @kev-ai/cli   # API + playground + CLI/MCP
pip install kev

See PUBLISH.md if you maintain the packages.

TypeScript

import { Choice, KevClient, Noul, Score } from "@kev-ai/sdk";

const client = new KevClient({ baseUrl: "http://127.0.0.1:3000" });

const res = await client.systemOne({
  state: "Package stuck in transit for a week. Tracking frozen.",
  questions: {
    topic: Choice("Which team?", {
      billing: "charges",
      shipping: "delivery / tracking",
      technical: "bugs",
    }),
    severity: Score("How urgent?", [
      "can wait",
      "this week",
      "today",
      "right now",
    ]),
    escalate: Noul("Page a human?"),
  },
});

console.log(res.answers.topic.choice, res.answers.escalate.noul);

LangChain / LlamaIndex (same package):

import { createLangChainKevTool } from "@kev-ai/sdk/langchain";
import { createLlamaIndexKevTool } from "@kev-ai/sdk/llamaindex";

Python

from kev import KevClient, Choice, Noul

client = KevClient(base_url="http://127.0.0.1:3000")
res = client.system_one(
    state="Charged twice. Furious.",
    questions={
        "topic": Choice("Which team?", {"billing": "charges", "technical": "bugs"}),
        "escalate": Noul("Escalate now?"),
    },
)
print(res.answers["topic"].choice, res.answers["escalate"].noul)

CLI & MCP

kev-server          # Decision API + playground
kev health
kev demo
kev ask --state "Charged twice" --trace
kev-mcp             # MCP stdio (Cursor / agents)

{
  "mcpServers": {
    "kev": {
      "command": "kev-mcp",
      "env": { "KEV_BASE_URL": "http://127.0.0.1:3000" }
    }
  }
}

From a clone (without global install):

pnpm --filter @kev-ai/cli exec kev demo
pnpm --filter @kev-ai/cli exec kev eval stability --trials 20
pnpm --filter @kev-ai/cli exec kev eval agreement --trials 20
pnpm --filter @kev-ai/cli exec kev eval multiq --multi-trials 5
pnpm bench:ops

How it works

Your app  ──POST /v1/systemone──►  Kev server
   ▲         state + questions         │
   │                                   ├─ validate (schema)
   │                                   ├─ decide (strategy)
   └──── typed answers + probs ────────┘
                                         │
              ┌──────────────────────────┼──────────────────────────┐
              ▼                          ▼                          ▼
        logprob readout           constrained JSON          parallel micro-score
        (when available)          (JSON mode)               (fallback)
              └──────────────────────────┬──────────────────────────┘
                                         ▼
                           Ollama · vLLM · OpenAI-compatible · mock

KEV_STRATEGY=auto picks the best path for your backend. Large choice sets cascade when needed. Optional: batch API, cache, rate limits, audit, traces.

Features

System One primitives — choice · score · noul with calibrated distributions
Self-host first — Docker-friendly, playground, OpenAPI
BYO model — mock · Ollama · any OpenAI-compatible endpoint
Production knobs — API keys, rate limits, LRU cache, audit log, OTel-style spans
Agent-ready — MCP server, LangChain / LlamaIndex adapters, images[] in state
Scale — cascadeChoice() for big taxonomies
Honest eval — held-out suite, ops latency/agreement/multi-Q, stability tests, published fixtures

Backends

`KEV_BACKEND`	When
`mock`	Offline demos & CI
`ollama`	Local GPU/CPU models
`openai`	OpenAI, vLLM, Groq, Together, Gemini OpenAI-compat, MLX, …

KEV_BACKEND=ollama
KEV_OLLAMA_MODEL=qwen3.5:9b

Model recipes: models/cards/

Clients & tools

	Install
TypeScript SDK	`npm i @kev-ai/sdk`
Server + playground	`npm i -g @kev-ai/server` → `kev-server`
CLI + MCP	`npm i -g @kev-ai/cli` → `kev` / `kev-mcp`
Python	`pip install kev`
LangChain / LlamaIndex	`@kev-ai/sdk/langchain` · `@kev-ai/sdk/llamaindex`
Recipes	`awesome-kev/`
Snippets	`.vscode/kev.code-snippets`

Publish notes: PUBLISH.md

Migrating from Jev

Point the base URL at NotJev : Kev. Re-tune thresholds on your labels.

export TYPESAFE_BASE_URL=http://127.0.0.1:3000
# or
export KEV_BASE_URL=http://127.0.0.1:3000

Full guide: MIGRATION.md

Project layout

apps/server        Decision API + playground host
apps/playground    Local workbench UI
apps/mcp           MCP server
packages/*         schema · core · backends · sdk · cli · eval · adapters
python/kev         Python client
benchmarks/        Held-out suite + fixtures
awesome-kev/       Recipes
models/            Model cards · calibration · dataset notes

Versioning

Kev 1.0 freezes the System One HTTP contract in GET /openapi.json.


MAJOR	Breaking wire changes
MINOR	Additive endpoints / fields
PATCH	Fixes and docs

See CHANGELOG.md.

License

Apache-2.0

Kev — decisions software can trust.
_{Ask for a label. Get a distribution.}

Name		Name	Last commit message	Last commit date
.cursor/rules		.cursor/rules
.github/workflows		.github/workflows
.vscode		.vscode
apps		apps
awesome-kev		awesome-kev
benchmarks		benchmarks
docker		docker
examples		examples
models		models
packages		packages
python		python
.env.example		.env.example
.gitignore		.gitignore
.npmrc		.npmrc
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
MIGRATION.md		MIGRATION.md
PUBLISH.md		PUBLISH.md
README.md		README.md
package.json		package.json
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
tsconfig.base.json		tsconfig.base.json
turbo.json		turbo.json

Navigation Menu

Repository files navigation

NotJev : Kev

Install

Docs & website

What NotJev : Kev is

Why NotJev : Kev

vs chat LLMs

What makes NotJev : Kev different

Benchmarks

Ops metrics (speed · structure · stability · batching)

Quick start

Playground

Use it in code

TypeScript

Python

CLI & MCP

How it works

Features

Backends

Clients & tools

Migrating from Jev

Project layout

Versioning

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

NotJev : Kev

Install

Docs & website

What NotJev : Kev is

Why NotJev : Kev

vs chat LLMs

What makes NotJev : Kev different

Benchmarks

Ops metrics (speed · structure · stability · batching)

Quick start

Playground

Use it in code

TypeScript

Python

CLI & MCP

How it works

Features

Backends

Clients & tools

Migrating from Jev

Project layout

Versioning

License

About

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages