Are the token and money savings measured or extrapolated?

Both exist and the README separates them. The A/B result is measured from provider-reported tokens: 99.3% fewer input tokens with accuracy up rather than down, 18/18 against 15/18. The slupek table and the daily, monthly and yearly money tables are modelled from a stated baseline plus list prices, and the author marks those as modelled. If you only trust one number, the author says trust the A/B one.

Which endpoint should I point it at?

Either, at identical published price. OpenRouter reports 32,000 tokens of context and a usage.cost field; the vendor endpoint reports 64,000, documents a 255 choice-option ceiling and 2 to 10 score levels, omits cost, and is behind a waitlist with keys issued in batches versus immediate access with credits. Measured from one machine, vendor warm p50 was 292-320ms and OpenRouter 325-371ms.

Does this replace my LLM for everything?

No, and the decision rule is stated bluntly: if the answer is one of N things you can list up front, Jev handles it; if it is prose, code or an open-ended set, use your LLM. jevskill plan enforces that split for free. The pitch is narrower than it first looks — you are replacing a yes/no gate and a shortlist step, not the model doing the work.

What if I have the wrong kind of key configured?

The skill detects the provider from the key shape, and you can override it with --provider or JEVSKILL_PROVIDER. With nothing stated, the first key found by name decides, vendor names first. Model names are translated in both directions because passing the OpenRouter id to the vendor endpoint is a 404 and passing jev-latest to OpenRouter is a 422. jevskill doctor checks the key, endpoint, latency and live cost.

How do I install jevskill and what does it need?

npx skills add lazniak/jevskill -g installs it for 20+ agents via the Agent Skills registry, with no account, no key and nothing to compile since the bundled scripts use only the Python standard library.

What does jevskill's A/B test actually measure?

Provider-reported tokens: 99.3% fewer input tokens with accuracy up, not down (18/18 versus 15/18). The token-savings table is explicitly modelled, but the 99.3% figure is measured.

How does batching questions change jevskill's speed?

8 questions in one call take 288 ms versus 3,564 ms asked separately, which is 12.4x faster and 4.03x fewer tokens because questions sharing one state run in parallel.

CommunityJev Agents & MCP

Jevskill

Agent Skill that sends large tool output and CI logs to Jev and hands your expensive model only the few lines that matter, shipped with a stdlib-only caller, a doctor command, and a plan mode that keeps open-ended work on your LLM.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Author: lazniak
Last verified: Oct 2026

lazniak/jevskill

User repository on GitHub

View on GitHub

Teach your coding agent to stop burning context. Jev (System One) via OpenRouter or TypeSafe: 325ms, 0.000013 USD per decision. A/B tested 99.3% fewer input tokens with accuracy up. Ships a reversible reduce and a ledger that learns when Jev pays off.

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Main language: Polish · English subtitles. The Skill is just the beginning. Next: ComputerUse. Your agent's context window is full of logs it didn't need to read. A Skill that makes Claude Code, Codex, DSH and 20+ other agents stop burning tokens on decisions — and start making them for $0.000013 in 325 ms. A/B tested: 99.3% fewer input tokens, and accuracy up, not down (18/18 vs 15/18). Jev is TypeSafe's System One decision model, and it is the whole engine here. Official model page: typesafe.ai · API docs.

Quoted from the project's README. Read the full README.

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Routing Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Three primitives: noul for yes or no, choice for one of N options, score for an ordered scale, each with a full distribution.
The 99.3% input-token reduction and the 18/18 versus 15/18 accuracy result are measured from provider-reported tokens.
Works through OpenRouter and TypeSafe's own endpoint, translating model names both ways since each rejects the other's id.
Bundled scripts use only the Python standard library, so npx skills add needs no build step, no account and no compiler.

Watch out

The context-growth tables and the per-session money tables are modelled from a stated baseline and published list prices, and the README explicitly says to trust the measured A/B token figure more than those numbers. The vendor endpoint returns no cost field, so the client computes cost from the published rate and marks the provenance as computed rather than recording zero and making every decision look free. jev-latest is a moving alias, so pin a version if you tuned a confidence threshold against one, and the vendor endpoint rejects session_id with a 400 because that field is an OpenRouter extension. Key detection works by variable name with vendor names first, so adding JEV_API_KEY next to an existing OpenRouter key moves traffic to the vendor unless you state --provider. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

muse-jev-playbook

Agent-oriented playbook for using Jev as a cheap, fast decision layer that triages, classifies, scores, and gates work before expensive browser, research, retry, or subagent steps.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

Bodila51/muse-jev-playbookPython

typesafe-skill-router

Opt-in Hermes plugin that routes each turn to one skill: a wide Jev Choice plus request checks gate a narrow fits round, injecting a single ignorable hint line.

Jev Agents & MCP#community#jevbest#catalog-import

Community

DECRUX9812/typesafe-skill-routerPython

tink-route

CLI that routes a task to one specialist skill from a Tink library using TypeSafe Jev, verifies the winner with tink mount, then prints it inline or as a mounted path.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

jon-devlapaz/tink-routePython

langchain-skill-router

LangChain deepagents middleware that replaces SkillsMiddleware with per-turn judging, loading only the SKILL.md files a turn needs so large catalogs stay out of the prompt.

Jev Agents & MCP#community#jevbest#catalog-import

Community

deyna256/langchain-skill-routerPython

jev-codex-router-skill

A Codex Skill that hands Jev the task summary plus the host's real model list, so Jev picks a model and reasoning effort for an independent subtask and Codex executes with it.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

455-dIAO/jev-codex-router-skillPython

jev-router

A Python CLI that asks Jev which of four coding agents should take a task, optionally waits for your approval, then opens that app or its website and pastes the prompt in.

Jev Agents & MCP#community#awesomejev#catalog-import

Community

AABBAASS1/jev-routerPython

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai