User repository on GitHub
Super fast Jev like server, model agnostic, working with any OpenAI compatible endpoint
- Forks
- 3
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Notjev
Zero-dependency client that turns any logprobs-capable chat endpoint into typed choice, boolean, and score decisions with abstention.
- Stars
- 22
- Language
- JavaScript
- License
- APACHE-2.0
- Published by
- Community
- Author
- 9pings
- Last verified
- Oct 2026
About this project
One token. A real probability. A decision you can trust (or refuse). Closed questions don’t need free-text generation. Present the options as A. B. C.…, ask for a single token with logprobs, keep only the letter mass, renormalise → you get a verdict + margin + tunable abstention. Works with any OpenAI-compatible endpoint that returns logprobs (vLLM, llama.cpp, Ollama, OpenAI…) Same model as your chat stack (true System 1) Speaks the Jev /v1/systemone wire contract. 26 options max. One HTTP request. maxtokens: 1.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Single-token readout with letter-mass renormalization into verdict, margin, portable band, coverage, and degraded flags.
- Twenty-six options max over one HTTP request with null for no verdict and prompt builder for offline diagnosis.
- Jev systemone wire support plus native decide route, CLI, service mode, and single-file binaries across platforms.
Watch out
Probabilities are ordered but uncalibrated and raw scores do not transfer across engines; thin states and large menus degrade accuracy. 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.
Closed questions need no prose: present options as letters, keep letter mass, and renormalize to verdict plus tunable abstention. When the margin is thin, escalate the same uncertain case to the same model in full chat mode.
Larger models hold up on rich two-option states and long menus while small fast ones suit filtering and triage with escalation. Menu size moves results more than model choice, so measure on your own states before trusting defaults.
Raw probabilities do not survive an engine change but the band does, with low, medium, high, and certain bands plus prior middles. Coverage flags when the model wanted something else and degraded marks low-mass answers.
Logprobs are the whole mechanism. notjev presents your options as letters and renormalises the letter mass from a single-token reply with max_tokens 1 and 26 options at most, so an endpoint that will not return logprobs cannot answer at all. Node 20 or later is the only hard requirement, and vLLM, llama.cpp, Ollama, or any OpenAI-compatible server works. Its published p50 numbers come from one RTX 5090: 19 ms for Phi-4-Q4_K_M and 23 ms for Qwen3-8B-Q4_K_M on llama-server.
npm install notjev on Node 20 or newer, with zero runtime dependencies. Single-file binaries are also published for every release tag across Linux, macOS and Windows on x64 and arm64, compiled with Bun from the same source. In those binaries the mcp subcommand refuses by name because the optional peers cannot be embedded; for MCP install the npm package alongside @modelcontextprotocol/sdk, and for the native backend add node-llama-cpp.
Yes, via notjev serve. It exposes /v1/systemone on the Jev wire contract so a typesafe-sdk works unchanged against any engine you configure, with answers keyed by question name and FastAPI-shaped 422s, plus a /v1/decide route carrying every Decision field. Setting NOTJEV_API_KEY bearer-guards the POST routes. Each answer carries a notjev block with margin, band, coverage and degraded, and its confidence is 1 minus H(p) over ln K, which is not comparable with Jev's numbers.
jev.prompt(q) builds the exact request string and sends nothing, which the README calls the first rung of every diagnosis. Two fields do most of the explaining: choice is null when there is no verdict, which is the absence of an answer rather than a fallback, and coverage is the option mass before renormalisation. Low coverage means the model wanted to say something you never offered, and degraded marks that it did.
More Jev resources like this
45GitHub stars
Drop-in OpenAI-compatible cache proxy for repeatable agent workflows: exact hits stay local, and with a TypeSafe key Jev judges whether a cached response is semantically reusable for the current request.
Jev Demos & Experiments#community#jevbest#catalog-import
33GitHub stars
Joke JavaScript helper that asks the Jev decision model whether a number is even and resolves to a boolean.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
wobsoriano/is-jevenJavaScript
wobsoriano/is-jevenJavaScript30GitHub stars
Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.
Jev Demos & Experiments#community#jevbest#catalog-import
24GitHub stars
Native macOS computer-use agent on Apple silicon that finds controls with local CoreML detection, OCR, and accessibility, then asks Jev to pick the next finite-choice action.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
jcpsimmons/jev-macos-loopJavaScript
jcpsimmons/jev-macos-loopJavaScript19GitHub stars
Six side-by-side web demos where Jev returns typed probabilistic decisions, plain code owns routing, and a switchable small model writes the words for Claude or Kimi tiers.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
mayank953/JevJavaScript
mayank953/JevJavaScript13GitHub stars
Chrome extension that adds a second Twitch chat column showing only messages Jev rates as matching your chosen intent, with every row linked to its probability distribution.
Jev Demos & Experiments#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.