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

jev-use

Claude Code, Codex and pi plugin that hands non-writing agent steps to Jev and escalates writing or unsure steps to the LLM.

PluginbetaMIT
View on GitHub ↗

Overview

jev-use makes the LLM and Jev collaborators inside coding agents: when a step just needs a fast decision, Jev answers it as a batched Noul, Choice or Score; when content needs writing, or Jev is unsure, a typed escalation contract hands the step back to the LLM. Questions that cannot be typed or need generation are rejected before the call, low-confidence answers come back flagged as priors, and a fail-open PreToolUse gate can only deny or ask, never silently allow.

One npx command wires Claude Code, Codex and pi, and a library with zero runtime dependencies on the judgment path is available on npm. It works with TypeSafe directly, OpenRouter or Vercel AI Gateway, and a mock backend runs keyless. Rerunnable demos include a directions task where Jev clicks and the LLM types, context compaction of 200 messages in 7 calls with 3 of 3 recall, a Pong game showing many more decisions per second than LLMs, and a shell-command gate that denies dangerous commands in about 230 ms. Listings report a measured p50 around 230 ms and about $0.02 per 1,000 judgments, with methodology and caveats in a results file.

What the decision model decides

TypeQuestionAnswers
Yes / noIs this shell command dangerous enough to deny or ask about?—
ChoiceWhich next step or UI target should the agent take?(inferred)—
Yes / noShould this message be kept during context compaction?(inferred)—

How it works

  1. Agent routes typed decision steps to jev-use
  2. Untypeable questions are rejected up front
  3. Jev answers in batch; low confidence flagged as priors
  4. Writing and unsure steps escalate to the LLM; gate denies or asks on risky tool calls

Strengths

  • Listings report p50 about 230 ms and about $0.02 per 1,000 judgments
  • Fail-open gate that can only deny or ask
  • Multiple providers and a keyless mock
  • Rerunnable benchmark scripts

Caveats

  • Measurements are the author's own
  • Judged state is sent to the configured provider

Getting started

npx -y jev-use install, then set TYPESAFE_API_KEY (or OPENROUTER_API_KEY / AI_GATEWAY_API_KEY) and run `npx -y jev-use doctor`.

At a glance

Type
Plugin
Status
beta
Language
JavaScript
Runtime
Agent plugin / MCP server / library
Has code
yes
License
MIT

Use cases and integrations

How curated lists describe it

  • MCP ecosystem: Claude Code / Codex / pi plugin (MCP server + library, native pi extension) that hands agent steps needing no text output to Jev as typed judgments — untypeable and generation-needing questions are rejected before the call, low-confidence answers come back flagged as priors, and a fail-open PreToolUse gate can only deny or ask.

    yibie/awesome-jev

  • Claude Code, Codex and pi plugin: Jev answers the batched typed questions an agent loop needs, and a typed escalation contract hands writing and low-confidence steps back to the LLM

    AnotiaWang/awesome-decision-models

  • Claude Code / Codex / pi plugin that hands agent steps needing no text output to Jev (TypeSafe's judgment model) — measured p50 ~230 ms and ~$0.02 per 1,000 judgments, with typed escalation back to the LLM

    logicrw/awesome-jev-projects

  • Claude Code / Codex / pi plugin that hands agent steps needing no text output…

    wh000wh000/awesome-jev-live

  • Claude Code / Codex / pi plugin where Jev answers batched noul, choice, and score questions and risk-checks tool calls, while a typed escalation contract hands writing and unsure steps back to the LLM.

    cobanov/awesome-jev

  • Claude Code / Codex / pi plugin that hands agent steps needing no text output to Jev (TypeSafe's judgment model) — measured p50 ~230 ms and ~$0.02 per 1,000 judgments, with typed escalation back to the LLM.

    hellogumbo/awesome-jev

  • Claude Code, Codex and pi plugin that hands the agent steps needing no written output to Jev and leaves the prose to the LLM.

    MrJev/awesome-jev

Frequently asked questions

What does Jev handle versus the LLM?

Jev answers steps needing no text output; the LLM writes content and takes over low-confidence steps.

Can the gate approve risky commands on its own?

No. The PreToolUse gate can only deny or ask, and fails open.

Can I try it without a key?

Yes, set JEV_BACKEND=mock for a keyless dry run.

Related projects

  • jev-mcp

    MCP server giving any agent Jev's classify, score, yes/no check and allow/confirm/block gating of risky tool calls.

    MCP serverbetaJavaScript
  • jevcore

    Jev decision layer for DeepSeek Harness, MCP hosts and plain Node, offline by default with disclosed egress and opt-in gates.

    MCP serverbetaTypeScript
  • jev-skill-router

    Local Python MCP service that keeps the skill catalog out of the agent's context and lets Jev pick which SKILL.md files to load.

    MCP serverbetaPython
  • jev-stack

    One-command npx installer for a set of Jev-powered agent tools: context compaction, browser agent, prompt enhancer and skill router.

    ProjectbetaJavaScript
  • Jevbridge

    Zero-dependency Node library with MCP and ACP adapters adding Jev decisions and gated computer use beside any LLM agent.

    MCP serverbetaTypeScript
  • jev-gate-mcp

    MCP gateway that trims noisy tool output and offers task classification and review advice, splitting work between local Laya and Jev.

    MCP serverprototypeTypeScript