Jev Skills

Agent skills for building with Jev, TypeSafe's model for typed decisions. OpenJEV support: Jev is built by TypeSafe. This fork keeps TypeSafe as the default and adds optional support for OpenJEV, a free community gateway to the same Jev model — set OPENJEV_API_KEY (or JEV_PROVIDER=openjev) to use it. Original project: https://github.com/laguagu/jev-skills by @laguagu.

The official skill covers the API. These cover the rest: where a decision fits, how to word the question, what code does with the answer, and how to show it works.

Install

npx skills add typesafe-ai/skills --skill typesafe-ai
npx skills add laguagu/jev-skills --skill jev-builder

Choose your agent when prompted, or add --agent codex. Claude Code plugin and other options.

Skills

Skill	Purpose
jev-builder	Setup, decision patterns, question wording and diagnosis, source-backed evidence
jev-evidence-eval	Accuracy, review rate, latency, and cost of a finished workflow

Use

Use jev-builder to add support-ticket routing. Include an unknown route and test the fallback.

Live calls need a TYPESAFE_API_KEY from the TypeSafe console. Installing a skill makes no API calls.

Examples

Clone the repository to run these. The skills work without them.

• Decisions: routing, ranking, tool selection, workflow control, risk scoring, and answer verification. Offline dry runs, plus tests for malformed answers, unknown choices, and spent retry budgets.
• Evidence: claim checking with source receipts, and a saved report whose confidence slider replays the policy without another API call.

Ecosystem

Reviewed on September 24, 2026 through public repositories and documentation. These are leads, not endorsements: apart from jegrep, nothing here was run for this kit, and most numbers are self-reported. Check license, maintenance, and failure paths before adopting any of it.

Learn

• TypeSafe docs and cookbooks: primitives, patterns, limits, and measured recipes.
• Confidence guide: designing an accept / review / fallback policy.
• LangChain: building a harness with Jev: worked routing and tool-screening code.
• Vercel launch notes: proposed places for Jev in an agent workflow; vendor figures.

SDKs, providers, and frameworks

• JavaScript and Python SDKs: official clients.
• System One Adapter: official stand-in for the Python client that answers with OpenAI, Anthropic, or Gemini. An LLM baseline on identical questions.
• Community clients for Java and Spring AI, Swift, Ruby, and Elixir.
• Vercel AI Gateway, OpenRouter, and Cloudflare Workers AI also serve Jev, each with its own key and request shape.
• LangChain TypeSafeClassifier: the three primitives behind a Runnable.
• Pydantic AI TypeSafeModel: agents with a structured output_type run on typesafe:jev-latest unchanged.
• LiteLLM pass-through: proxy the endpoint with logging and spend tracking.

Agent skills and MCP

• typesafe-ai/skills: the official skill; keep it installed from upstream.
• dbreunig/building-with-jev-skill: question design, composition, and diagnosis for jev-1.13. No license file at review.
• altryne/jevify: steers an agent to hand bulk semantic checks to Jev during ordinary work, with standard-library Python helpers.
• wuyoscar/jev-skill: six skills for triage, documents, eval, UI, and simulation, with recorded requests and responses.
• kerpopule/hermes-jev-skills: routing, memory, compaction, skill selection, and computer and browser use, with fail-open defaults and a shadow mode.
• AutoJev skills: task and model routing, tool checks, completion review.
• jkudish/jev-mcp: ten task-shaped MCP tools (verify, screen, rerank, gate…) that fail closed on malformed answers. itsmostafa/typesafe-mcp exposes one general evaluate tool instead.
• tamaratran/fast-jev-compaction and jev-pruner: Claude Code plugins that choose which tool results survive compaction and trim long Bash output. Read their hooks before enabling.

Open and local models

Same request shape, different models. None of these is Jev, so re-measure any threshold you carry over.

• jaredpalmer/kev: 0.8B to 9B decision models on Qwen3.5 with training code. Serves the System One API, so the official Python SDK can point at it.
• TheoLeeCJ/SemIf-OpenJev: reads option probabilities from open models, including a WebGPU demo that runs in the browser.
• featherless-ai/simple-jev: choices, scores, and yes probabilities from Hugging Face model logits, with a keyless demo API.
• zhengxuyu/litjev: the same idea on Qwen checkpoints. Useful for understanding the shape of the API.

Browser and computer use

• browser-use/jev-ultrafast: picks an operation and its target element in one request over numbered page elements.
• ndrezn/ts-browser-agent: the same idea on langchain-typesafe.
• realZachi/typesafe-adblock: a small extension that asks whether a DOM element is an ad. A minimal per-element decision loop.

Developer tools

• can1357/jegrep: semantic grep that scores files by probability, without an embedding index. Run for this kit: it found every labelled file for its own English benchmark queries, but under half when the same questions were asked in Finnish, because its candidates come from a keyword scan (benchmark).
• andududu/jeview: a local gateway that records every call and draws it live. Handy when a question misbehaves.
• sutro-sh/jev-align: finds uncertain rows, asks you to label them, and proposes a better question definition with GEPA.
• valentynkit/jev-commit: pre-commit hook judging whether the message matches the diff.

Applications

• devagrawal09/jev-review: staged code review. A Noul risk matrix, then evidence selection and severity scores, with thresholds in code.
• kyotofin/tax-doc-classifier: a Choice over 261 IRS forms per page, scored strictly with a confidence gate and a reproducible eval.
• fazlerocks/jevmail: inbox triage into reply, update, promotional, sales, and spam trays, with an urgency score.
• devanshbatham/commit-miner: classifies commit diffs into fixes, security changes, and CWEs.

Directories

• cobanov/awesome-jev: favours public source and reproducible evidence.
• logicrw/awesome-jev-projects: every entry pinned to a commit and a decision point.
• yibie/awesome-jev and AbdelStark/awesome-typesafe-jev: categorised field guides.
• github.com/topics/jev: everything tagged, unfiltered.

Several lists index the same launch-week repositories, and one author can publish many at once. Prefer the original artifact, and one repository that shows a real Jev call over ten that describe one. The resource guide is the version your agent reads.

Independent community project. MIT license.