Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

jev

CPU-only local Jev-style decision service plus an oh-my-pi extension that retargets the agent's judge provider to it.

Pluginprototype
View on GitHub ↗

Overview

jev-local is a loopback-only, CPU-only service that runs the open Jev-Style-0.8B-Decision-v3 GGUF model and exposes the oh-my-pi systemone and alpha/decisions contract over HTTP, together with an oh-my-pi extension that points the agent's typesafe judge provider at it. Judgments then need no GPU, no cloud calls and no API key.

The scorer reads decisions directly from llama.cpp as the difference between the logits for yes and no at each arrow slot, then applies a softmax with a fitted calibration temperature, never generating text. A systemd user unit keeps the service running across logouts and reboots, with the installer setting a measured memory limit. The README explains its choice of port 61021, just above Linux's default ephemeral port range, so the kernel cannot claim it for outbound connections. The 505 MiB model is fetched separately rather than stored in the repo.

What the decision model decides

TypeQuestionAnswers
Yes / noJudge question from the oh-my-pi agent(inferred)yes · no
ChoiceChoice question from the agent(inferred)—

How it works

  1. Fetch the 0.8B GGUF model
  2. Install the systemd user service on 127.0.0.1:61021
  3. Install the oh-my-pi extension to retarget the judge provider
  4. Agent judgments are scored locally from logits

Strengths

  • No GPU, cloud calls or API key
  • Bound to loopback only
  • Survives reboots via systemd user unit

Caveats

  • Small 0.8B model will be less accurate than hosted Jev
  • Specific to the oh-my-pi agent

At a glance

Type
Plugin
Status
prototype
Language
C++, TypeScript
Runtime
local service, Pi extension
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

What does this replace?

The oh-my-pi agent's hosted TypeSafe judge, with a local CPU-only model.

Does it generate text?

No. It reads yes/no logit differences at fixed slots and calibrates them.

Why port 61021?

It sits above Linux's default ephemeral port range so the kernel cannot take it.

Related projects

  • system-one

    Provider-neutral TypeScript runtime: write choice, noul and score once and run them on Jev, Reflex or any System One endpoint.

    ServicebetaTypeScript
  • decidealot

    Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.

    MCP serverbetaPython
  • instinct-ai

    Open-source dual-brain runtime that answers typed decisions with a fast local engine or the Jev API before falling back to LLMs.

    ServicebetaPython, TypeScript
  • jev-mcp

    MCP server exposing Jev-style decision models served locally by Ollama's /v1/systemone as classify, yes/no and score tools.

    MCP serverbetaTypeScript
  • noulxp

    NoulXP: open standard and reference engine for packaging calibrated System One decision models so one runtime can run them all.

    ProjectbetaPython
  • openpave-jev

    PAVE skill and CLI for typed Choice, Score and Noul decisions via local Laya, any /v1/systemone server, TypeSafe Jev or an EPM gateway.

    Agent skillbetaTypeScript