Jev Skills

A collection of agent skills, repositories and examples for TypeSafe Jev. Jev answers typed questions about text: choose an option, score it, or return the probability of yes. Your code decides what happens next.

🧩 Skills · 📦 Install · 🛠️ What you can build · 🚀 Projects · 🔌 SDKs & integrations · 🧪 Examples · 📚 Learn · 🏠 Open models · 📊 Evaluations

New to Jev? Read the quickstart, install the two starting skills, or try an offline example.

🧩 Skills

📁 In this repo

🌍 Official and community skills

📦 Install

Start with the official API skill and this repo's builder skill:

bunx --bun skills add typesafe-ai/skills --skill typesafe-ai
bunx --bun skills add laguagu/jev-skills --skill jev-builder

Choose your agent when prompted, or add --agent codex. Add -g for a personal installation. Replace jev-builder with jev-evidence-eval or jev-curator for either companion. Plugin and other installation options.

🛠️ What you can build

flowchart LR
    accTitle: A Jev decision in an application
    accDescr: Text and typed questions go to Jev. The application uses its answers to act, review or fall back.
    S[Text + typed questions] --> J["Jev<br/>Answers + probabilities"]
    J --> P["Your code<br/>Act, review or fall back"]

Read the official cookbooks for worked recipes. Code handles permissions, thresholds and fallback behavior; measure decisions on your own cases.

🚀 Projects

🤖 Coding agents & MCP · 🌐 Browser & desktop agents · 🔎 Search & data · 🧰 Developer tools · 📱 Applications · 🎮 Games & simulation

Descriptions summarize each project's docs and source; a listing does not mean it was run here. Sources and repository status checked September 30, 2026.

🤖 Coding agents & MCP

• jev-mcp — MCP tools for verification, screening, reranking and decision gates.
• system-one-connector — A general evaluate MCP tool; formerly typesafe-mcp.
• jev-router — Model routing for Claude Code and Codex.
• fast-jev-compaction — Select which tool results survive Claude Code compaction.
• jev-pruner — Trim long Bash output before it reaches the agent.
• winnow — Claude Code hook that replaces unneeded tool-result sections with expandable stubs.
• jev-skill-router — Claude Code hook for skill suggestions; logs decisions in shadow mode by default.
• SkillRanker — Rust CLI that ranks installed agent skills for the next step from the live session.
• eran-broder/jev-skills — Claude Code and Codex hooks that load only the skills Jev selects for the current step.
• Canny — Claude Code and Codex hooks that refuse "done" until a check has passed; Jev's rule judgments add notes but never block.
• jev-guard — Risk-scores each tool call before it runs: an auto mode for several coding agents.
• pi-jev — Pi coding agent extension that gates bash, write and edit calls and adds a jev_ask tool.
• Foreman — Supervises Codex or OpenCode workers, judging whether a job is complete, tested enough or needs a person.

🌐 Browser & desktop agents

• jev-ultrafast — Browser agent that selects an action and a page element together.
• Jev for Chrome — Extension that drives your current tab, choosing each click, keystroke or dropdown value.
• ts-browser-agent — Browser action selection through LangChain.
• agent-desktop — Desktop automation over accessibility trees; its jev-desktop skill picks each operation and its target in one request.
• typesafe-computer-use — Mac automation with Jev choosing actions and a text model filling free text. Beta.
• mobile-jev — Android action selection over Mobilerun screen controls.
• Jev Social — Read-only research on Instagram, TikTok and LinkedIn, with Jev choosing each next browser step.

🔎 Search & data

• jegrep — Semantic code search without an embedding index.
• jevgrep — Repository questions answered with relevant files and verbatim excerpts; includes an agent skill.
• blink — Codebase search using parallel walkers guided by Jev. No repository license at review.
• Jev Search — Web search in which Jev picks the sources, time range and query, then scores each result.
• jev-reranker — Python library for reranking and relevance filtering.
• Vector Graph RAG — Graph RAG on Milvus with an optional Jev reranker for multi-hop relations.
• JevPDF — Search a PDF by meaning and highlight matching lines.
• neo4jev — Navigate a Neo4j graph by choosing relationships at each hop.
• sqlite3-jev — Ask Choice, Score and Noul questions from SQL.
• jevQL — A jev() function family for plain PostgreSQL, usable from a CLI, an MCP node or SDKs.
• truffler — Intent search for Rails with stored labels and query-time reranking.
• jev-curate — Filter and score large Parquet or JSONL training datasets. Beta.

🧰 Developer tools

• jev-review — Code review in stages: risk, supporting evidence and severity.
• Supercov — Scores code quality and security risk with Jev and turns coverage gaps into tasks for a coding agent.
• jeview — Local gateway for inspecting requests and answers as they happen.
• jev-align — Label uncertain examples and refine question definitions with GEPA.
• jevcal — Calibrate thresholds and check for drift after model changes.
• jevals — Agent-trace evaluations and guardrails.
• SemDecide — Typed decisions in Unix pipelines and CI: pipe text in, get a probability and an exit code.
• jev-commit — Pre-commit checks against the staged diff and commit message.
• jev.nvim — Rank functions in a Neovim buffer against a natural-language question.
• commit-miner — Classify commit diffs into fixes, security changes and CWEs. No repository license at review.
• decision-gate — Shared rate limits, spend limits and answer caching for decision loops.

📱 Applications

• jev-cookbook — Runnable Node recipes for support triage, tagging, deduplication, search and more.
• docjev — Classify documents and split document packets using parsed page text.
• tax-doc-classifier — Classify IRS forms page by page, with an evaluation harness.
• jevmail — Inbox triage with categories and urgency scores.
• jev-usecases — Workflow examples that map decisions to automatic action, confirmation or review.
• Notra — Content platform whose agent uses Jev to choose a fast, everyday or deep model per turn.
• OpenWork — Open-source Claude Cowork alternative; its eval testkit can have Jev pick which checked-in checks cover a verification request. Experimental.
• Shapeshift — A text box that turns into an event card, checklist, timer or poll as you type; Jev classifies the intent when a key is set.
• unclutter — Browser extension that hides nonessential page elements and keeps the result as reusable rules.
• typesafe-adblock — Browser extension demonstrating per-element ad detection.
• Jev Moderation Bot — Discord bot that deletes spam and scam links and escalates repeat offences.
• Kill My Idea — Scores a startup idea on eight questions and answers kill, fix or ship. No repository license at review.
• jev-trader — Monad market-making demo with buy-or-sell decisions; simulates fills without a wallet key.
• Prism — Solana liquidity agent whose rule-based rebalancing also asks Jev about toxic flow, holding and deposit shape.

🎮 Games & simulation

• typesafe-mario — Choose Mario actions from structured emulator state. No repository license at review.
• OneVOneJev — Browser arena with a Jev-controlled opponent. No repository license at review.
• tsai-sc — Jev plays the first StarCraft shareware mission by keyboard and mouse, with recorded action probabilities.
• Trick Jev — Browser game: change Jev's answer one word at a time.
• jev-drone — Camera-only quadrotor in MuJoCo: Jev reads the situation a few times a second while code flies.

🔌 SDKs & integrations

🧑‍💻 Clients

Links to the typesafe-ai organization are official; the others are community clients.

🏗️ Providers & frameworks

• Vercel AI Gateway · AI SDK provider
• OpenRouter — Hosted decisions endpoint.
• Cloudflare Workers AI — Jev through a Workers binding.
• LangChain — Typed classification in a Runnable.
• Pydantic AI — Structured agent outputs through TypeSafe.
• AG2 — Decision agents for yes/no questions, label selection and rubric scores.
• eve — Vercel's agent framework; its auto model selection asks Jev by default.
• json-render — Experimental, unreleased UI composition in which Jev chooses components from your catalog.
• DSPy — Decision types and optimization examples.
• LanceDB · Milvus — Reranker integrations.
• Bifrost · LiteLLM — Gateway integrations.
• DeepEval — Evaluation metrics from typed questions.
• Opik — Traces for System One calls.
• Hono — Experimental router that matches HTTP requests to plain-language descriptions.
• Home Assistant — Typed questions about your home, exposed as entities for automations.

🧪 Examples

Small examples included in this repo, with synthetic inputs and offline checks.

Try a dry run with Node.js 22+; no API key or dependency install needed:

git clone https://github.com/laguagu/jev-skills.git
cd jev-skills
node examples/decisions/run.mjs routing --dry-run

📚 Learn

• TypeSafe documentation — Concepts, primitives and API reference.
• Jev with coding agents — How the official skill helps an agent write code that uses Jev.
• Official cookbooks — Worked recipes for routing, retrieval, verification and interactive apps.
• Use-case map — Find where a typed decision fits.
• Confidence guide — Decide when to accept, review or fall back.
• Building a harness with Jev — Routing and tool screening with LangChain.
• Search with Jev and Milvus — Nine notebooks: reranking, context filtering, query routing, cache reuse and knowing when to stop searching.
• The Top 20 Jev Skills — Charlie Hills' tour of agents, tools and demos.

🗂️ More collections

• yibie/awesome-jev — Projects grouped by use case. No repository license at review.
• kraayenjon/awesome-jev — Builds, SDKs, guides and community resources.
• heyjunpenn/awesome-jev — A large catalog of open-source projects with a searchable website.
• Anil-matcha/awesome-jev-by-typesafe — Use cases, prompt patterns and starter code.
• cobanov/awesome-jev — Projects with public source and reproducible evidence.
• logicrw/awesome-jev-projects — Entries tied to source commits and decision points.
• AbdelStark/awesome-typesafe-jev — Field guide with SDKs, runnable projects and an agent skill.
• OmniJev/awesome-jev-gallery — Papers, open models and evaluations.

Thanks to these collections and Charlie Hills for discovery leads. Descriptions here are written from the linked primary sources.

🏠 Open models

Alternatives for exploring the decision-model interface locally. They are separate models with their own behavior; their probabilities and thresholds are not interchangeable with Jev's. The license shown is the repository's; check the model weights separately.

• kev — Open decision models, training code and a System One API server. Apache-2.0
• Laya — Local decision models for the three primitives. Apache-2.0
• Ollaya — Pull and serve Laya, decider and other open decision models behind a TypeSafe-compatible API, plus an MCP server. Apache-2.0
• Laya-MLX — Runs Laya checkpoints natively on Apple Silicon. Apache-2.0
• SemIf-OpenJev — Typed decisions from open models, including a browser demo. MIT
• simple-jev — Choices, scores and yes probabilities from model logits. Apache-2.0
• decider — Qwen3.5 fine-tunes that answer typed questions in one forward pass. Apache-2.0
• Von — A non-autoregressive decision model for choices, probabilities and ordinal scores. Apache-2.0
• NanoJev — A 0.6B parallel decision model with its training pipeline and game demos. MIT
• jevos — Yes/no probabilities on a laptop CPU. MIT
• Jevlike — Train a small model that picks among a changing list of text options. MIT
• litjev — Decision API experiments on Qwen checkpoints. Apache-2.0
• AnyJev — Calibration heads for open models, served through vLLM. Apache-2.0

Model weights

Hugging Face downloads for the open models above. Each is Apache-2.0 on its model card.

Discover models & datasets

• System One models on Hugging Face — Browse model cards tagged system-one for weights, usage and licenses.
• System One datasets on Hugging Face — Browse datasets tagged system-one for training and evaluation; check each card's provenance, splits and license.
• mindchain's System 1 collection — A community collection of decision models and datasets.

Skills for open models

Skills for the open alternatives below. Install from upstream; thresholds and probabilities are specific to each model, so measure on your own cases.

bunx --bun skills add ollaya-dev/ollaya --skill ollaya-decisions

📊 Evaluations

• System One Mosaic Benchmark (S1MB) — English Choice, Noul and Score comparisons, with links to evaluation code, datasets and a leaderboard; the article explains scoring and training-overlap limits.
• Jev Decision Index — A community leaderboard comparing open Jev reproductions with Jev, with per-model results, methodology and ecosystem news.
• TypeSafe workflow evals — The provider's workflow measurements.
• S1Rank — Retrieval benchmarks, calibration and raw responses.
• ASSAY-001 — Pre-registered calibration and type-safety check on two intent datasets, with every raw response. No repository license at review.
• sysone-bench — Decision models compared on matching inputs.
• jev-spam-eval — Email classification against a trained baseline.
• JevBench — Decision-model comparisons.

Hugging Face discovery and evaluation links checked October 4, 2026.

🔄 Keep the collection current

Ask your agent:

Use jev-curator to find useful new Jev repos and skills, check the original sources, and update this README.

In a checkout, the agent can read the skill directly. Contributions welcome: link the original project and explain its use in one sentence.

Independent community collection. MIT license; linked projects have their own licenses.