engineering / skill

jev-use

A plugin and library that hands agent steps needing no text output to Jev, with typed escalation back to the LLM, for Claude Code, Codex and pi.

Repository owner shitianfang

Not rated yet
0 comments

Like 0 Save

Published 11 Oct 2026Listing updated 12 Oct 2026

At a glance

Purpose: A plugin and library that hands agent steps needing no text output to Jev, with typed escalation back to the LLM, for Claude Code, Codex and pi.
Runs in: Local agent with shell, MCP server
Operating systems: Not stated
Needs: Runtime: Not stated; Needs internet: Yes; Needs accounts or API keys beyond your agent: Yes
Files and data: Reads files: Not stated; Writes files: Not stated; Sends data to other services: Yes
Tool type: Skill
Delivery: From GitHub
Version: Latest on the main branch
Access: Code on GitHub
Compatibility: According to the README, it works with Claude Code, Codex and pi, and its install command wires whichever of them it finds. The README names TypeSafe direct, OpenRouter and Vercel AI Gateway as Jev providers, and shows library use through npm.
External costs: The repository is MIT-licensed. Real use needs your own key for one of the three Jev providers the README names (TypeSafe direct, OpenRouter or Vercel AI Gateway), and those providers may charge for usage. The README points to bench/RESULTS.md for its numbers. Your coding agent or model provider may also charge for usage.
Permissions: The plugin acts through your coding agent with the access you give it. The README shows installation through npx -y jev-use install and library installation through npm i jev-use. It says the install command wires Claude Code, Codex and pi, whichever it finds, but does not describe the installation's write locations. The README instructs users to supply TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY for provider use. It describes plugin forms with a PreToolUse gate and a demo that gates shell commands. The README lists two MCP tools and describes src/redact.ts as stripping credentials from a gated action before transmission. The top-level file list includes scripts/, but the README does not describe that folder's permissions.
Usage terms: MIT License, as stated in the repository. See its LICENSE file for the exact terms.

Code on GitHub

shitianfang/jev-use

MIT

JavaScript60 stars14 forksLast pushed 22 Sept 2026 UTC

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

Topics: agent agent-skills ai-agents claude-code claude-code-hooks claude-code-plugin claude-skills codex

Public GitHub data fetched 13 Oct 2026 UTC · 0 days old

Skill folders in this repository (1)

Folders that contain a SKILL.md file, from GitHub's file list on 13 Oct 2026 UTC.

skills/jev-use

View on GitHub ↗

What it does

jev-use is a plugin and library for Claude Code, Codex and pi. According to the README, it hands the tasks that need no text output to Jev, which the GitHub description calls TypeSafe's judgment model. The README says that when content needs to be written the LLM takes over, and when a step only needs a decision Jev executes it.

The README's library example asks Jev three kinds of question: check, pick and rate. The example result for one question holds the answer, a confidence and an escalate flag. The README says anything Jev can't or shouldn't decide comes back with escalate set to true and a typed reason. It shows four demos, such as gating shell commands and context compaction, and points to bench/RESULTS.md for numbers, methodology and caveats.

The repository holds the TypeScript source, including the code for the two MCP tools, a routing skill at skills/jev-use, plugin forms for Claude Code and Codex with that skill and a PreToolUse gate (harness/claude-code and harness/codex), benchmarks and demo scripts. The README lists what each source file does.

Who it is for

People who use Claude Code, Codex or pi and want to hand quick decision steps to Jev.

Example tasks

Run the install command to wire Claude Code, Codex and pi, whichever it finds.
Run a keyless dry run with the mock backend.
Check the wiring with the doctor command.
Ask Jev check, pick and rate questions about a run's state from code.
Rerun one of the demo scripts in bench/examples.

Getting started

The README's install command is npx -y jev-use install, which it says wires Claude Code, Codex and pi, whichever it finds. Set one key in the environment your agent runs in: TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. Use JEV_BACKEND=mock for a keyless dry run. The README says npx -y jev-use doctor checks the wiring. For library use it shows npm i jev-use. The README has the exact commands.

Limitations

Real use needs a key for one of the three Jev providers the README names (TypeSafe direct, OpenRouter or Vercel AI Gateway). The README's keyless dry run uses the mock backend. It presents the tool for Claude Code, Codex and pi. The README presents speed, token and quality claims alongside its demos. It points to bench/RESULTS.md for numbers, methodology, variance and caveats, and to docs/evidence.md for third-party measurements. The repository can change at any time; check its README before you install.

Data practices

According to the README, judged state goes to the provider you configure, while JEV_BACKEND=mock stays local. The README describes src/redact.ts as stripping credentials from a gated action before transmission and src/dispatch.ts as providing pre-call routing. The README does not mention telemetry.