jev-router
Route Claude Code tasks to a suitable lower-cost model from a defined set of candidates.
Find Jev-related repositories, agent skills and community experiments. Choose a use case to explore the code behind a project.
Route Claude Code tasks to a suitable lower-cost model from a defined set of candidates.
Choose a skill for an agent, with explicit no-skill and review paths when the match is uncertain.
A local proxy that selects a Claude model and reasoning effort for each message.
A LiteLLM-based router that uses Jev to choose a model for each request.
Assess shell commands for destructive actions, data transfer, remote execution and weakened security before an agent acts.
A guard layer for coding agents that evaluates prompt injection and potentially dangerous operations.
Inspect dependency source and build files, then examine suspicious files and lines more closely.
Scan changes for secrets and destructive commands before a Git commit or push.
Watch coding-agent changes and flag potential violations of project rules for review.
Evaluate pull-request files against 31 code-quality questions, then pass findings to a writing model.
Kyle Jeong’s browser-agent demonstration: Jev chooses an action; Stagehand carries it out.
A browser-use agent that lets Jev select the next action and uses a generative model when text is needed.
A browser extension that identifies distracting page elements and removes them.
A Chrome extension that turns ad detection into a series of yes-or-no judgments.
Let coding agents operate a real Chrome browser, with Jev helping choose actions.
Check paragraphs for vague phrasing, repetitive endings and other writing patterns using focused judgments.
Evaluate note.com articles across eight dimensions and combine them into a content-quality score.
Use Claude to locate cited passages and Jev to assess support, with a person making the final call.
A demonstration of categorizing email, judging priority, detecting spam and deciding whether a reply is needed.
A marketing-analytics platform exploring Jev for brand-visibility classification.
Classify incoming messages as invoices or general email and send them to the corresponding workflow.
Filter synthetic datasets using condition judgments and confidence, saving accepted and rejected records separately.
A resume-screening experiment with editable criteria and reusable scores. Human review remains essential.
An issue-triage dashboard covering type, severity, urgency, duplicates and next steps, with recorded human corrections.
Keep useful tool calls and results in a coding session while removing irrelevant context without summarizing it.
A proxy layer for Claude Code and Codex that evaluates which conversation history is still needed.
Check whether a scheduled agent wake-up is necessary before waking a larger model.
A block-by-block trading experiment. The shared demonstration shows dry-run mode; it is not evidence of profitability.
A Home Assistant integration that answers questions about home state with choices, scores and probabilities.
An experimental painter that selects pixel colors in parallel and uses confidence to vary brush width.
An open-source Pac-Man experiment where Jev chooses which direction to move next.
A Pong experiment comparing model-controlled decisions as the ball moves.
A robotic-arm workbench that selects joint actions from candidates and includes a spending limit.
Fit confidence thresholds to labeled examples and estimate fallback rates, with checks for model changes.
A reported test of 33,047 entries and 164 queries where Jev-only reranking did not outperform vector retrieval.
An independent 0.6B-parameter decision-model implementation with training code, weights and data.
An independent non-autoregressive decision-model project supporting three question types.
A small independent decision-model experiment based on Qwen2.5-0.5B, designed to train and run on a MacBook.
Try another search, category or link type.
Use the repository documentation for installation. For the TypeSafe integration skill and a first decision workflow, follow the resources below.