Are the cost and accuracy numbers proven?

They are calibration observations, not guarantees: the README reports a fraction of a cent per judged node and clean risk-band separation on a 12-node labeled fixture, while warning all thresholds are proposed and uncalibrated.

How is this different from a one-shot AI reviewer?

One-shot reviewers read a diff once and stop. This pipe expands each lead across callers, co-changed files, and configs, judges every node cheaply, and verifies citations before anything becomes a finding.

Can a hallucinated finding become a bug report?

Not through the designed path: the verifier reads each cited file and line off disk, and anything with fabricated or missing citations ships as an unverified lead, never a bug.

Will it leak my secrets to the model?

Credential-shaped values are redacted before network calls, deny-listed secret files are never read, and env files are skipped. Seed, expand, and verify still operate on raw text locally.

How do I install it and try it for free?

It needs Node 20 or newer, and npm install pulls the one dependency, ai, used by jev-judge for the Jev call. The first run can cost nothing: ./bin/crawl --repo ./examples/todo-app --dry-run --budget 8 stubs judgments, so it needs no key. npm test runs the 46 mocked tests across 9 files.

What are the stages of a crawl?

jev-seed emits starting leads; jev-expand follows context cues mechanically with no model calls; jev-judge makes one Jev call per node for a verdict, a bug-likely signal, a risk score and an artifact flag; jev-verify checks every cited file and line against disk; jev-report renders findings; jev-verdict records human outcomes; and crawl is the orchestrator with a budget, depth cap and stop rules.

Can I use the verifier on its own?

Yes. jev-verify is deliberately dependency-light: Node 20+, one shared helper file, no Jev calls, no key and no crawler imports. It takes node and judgment records on stdin from any pipeline and writes findings on stdout; exit 0 means the claim cites real code, and exit 1 means it did not survive verification.

CommunityJev Integrations & Adapters

Jev Crawlers

Unix-style bug-discovery pipeline: small tools pass JSON lines through seed, expand, judge, verify, and report stages, spending one cheap Jev judgment per code node to chase leads across callers and configs.

See highlights

Open github.com

Stars: 1
Language: JavaScript
License: MIT
Published by: Community
Author: Thingscorp
Last verified: Oct 2026

Thingscorp/jev-crawlers

Organization repository on GitHub

View on GitHub

Jev learns your repo's decision norms, then adversarially judges past decisions against them. Unix-style primitives (seed, expand, judge, verify, report, norms) with per-node typed judgments from typesafe-ai/jev.

Forks: 1
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Jev learns your repo's decision norms, then adversarially judges past decisions against them — as small unix tools for bug-discovery: one tool, one job, JSON lines on stdin/stdout, composed with pipes.

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Seed emits leads from diffs, TODO and FIXME comments, and risky patterns; recorded false positives suppress repeats.
Expand follows callers, git co-change history, and config references mechanically with no model calls.
Judge returns verdict, bug-likelihood rank, 0-3 risk score, and artifact presence; routing follows risk bands.
Verify grounds every file and line citation against disk; fabricated evidence demotes to an unverified lead.

Watch out

Self-described experiment with a small labeled calibration fixture, proposed uncalibrated thresholds, and verification that checks grounding rather than execution. 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

dsh-opencli

DeepSeek Harness plugin that gives agents hands in the user's own logged-in browser: 176 site adapters with 200+ one-shot commands plus generic page primitives, with write actions behind an approval gate.

Jev Integrations & Adapters#community#github

Community

IKEASven69/dsh-opencliJavaScript

brasileirao-jev

Claude Cowork plugin running reproducible Brasileirao Serie A probability experiments through Jev over Vercel AI Gateway, with Claude only fetching the table and presenting raw numbers.

Jev Integrations & Adapters#community#github

Community

andretestanalyst/brasileirao-jevJavaScript

dsh-jev

DeepSeek Harness community plugin that classifies each turn's newest user message through the systemone endpoint and injects the verdict as tagged runtime context for tone and priority.

Jev Integrations & Adapters#community#github

Community

Excalibur9527/dsh-jevJavaScript

jev-gateway

Local HTTP service exposing Jev through the Vercel AI Gateway: projects POST state plus typed questions and get Jev answers with provider metadata, without holding a TypeSafe API key.

Jev Integrations & Adapters#community#github

Community

nitinnat/jev-gatewayJavaScript

ai-backends

146

A self-hostable API server that fronts text, vision and data endpoints across Ollama, LM Studio, LlamaCpp and a long list of hosted providers, and also runs tool-using agents with sessions. Jev appears separately as a decision provider for typed choice, score and noul questions.

Jev Integrations & Adapters#community#jevbest

Community

donvito/ai-backendsHTML

ST-jeved

SillyTavern extension that sends chat text to Jev sensors and fires one rule action per turn such as nudging, rerolling, editing lists, or running scripts.

Jev Integrations & Adapters#community#jevbest#catalog-import

Community

mossyfield/ST-jevedJavaScript

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai