User repository on GitHub
Guardrails for LLM apps in one API call. Prompt injection, jailbreaks, leaks, unsafe content. Built on TypeSafe Jev. MIT.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Tools & CLIs
Jevegis
Jevegis wraps Jev into a single guardrail call for LLM apps: one request covering prompt injection, jailbreaks, credential and PII leaks, unauthorized actions, malicious code and indirect injection in retrieved documents, plus Trust & Safety moderation. It returns probabilities rather than booleans.
- Stars
- 2
- Language
- TypeScript
- License
- MIT
- Published by
- Community
- Author
- 0xArx
- Last verified
- Oct 2026
About this project
Open source. MIT licensed. Live at SDK/CLI: Guardrails for LLM apps in one API call. Security (prompt injection, jailbreaks, credential/PII leaks, unauthorized actions, malicious code, indirect injection in retrieved documents) and Trust & Safety moderation, both powered by TypeSafe's Jev typed-judgment model: calibrated probabilities, no generated text, sub-second calls.
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
Highlights
- The detection engine is one declarative checks array, and each check knows how to phrase itself per target: input, output, document or content.
- evaluate() builds a single Jev request, applies thresholds, and returns verdict, reasons, flags, category and severity in one round trip.
- Requests parse from text, direction or OpenAI-style messages, where the last message is judged and the rest is treated as context.
- npm run eval replays every labeled case against the live engine and reports verdict accuracy, per-flag misses, p50 and p95 latency, and cost.
Watch out
This is a hosted service first and a library second: the authenticated scan and moderate routes sit behind bearer keys with Postgres rate limiting and usage logging, and running it yourself means provisioning a Supabase project, applying the schema, and supplying a TypeSafe key plus the Supabase URL and service-role key. Usage logging is described as best-effort and never fails a scan, so an audit trail is not guaranteed. The eval suite is 42 hand-labeled cases scored against the live engine, useful as a smoke test rather than as evidence about adversarial traffic, and the SDK and CLI live in a separate repository. Auth runs through Supabase magic links. 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.
The engine holds one declarative checks array, and each check knows how to phrase itself for the target it is inspecting: input, output, a retrieved document, or generated content. evaluate() builds a single Jev request, applies the thresholds and returns a verdict with reasons, flags, category and severity. So the breadth comes from how many checks you declare, and you can narrow or widen that set per request through the checks and thresholds fields.
It lets you set your own bar rather than inherit someone else's block threshold, and the returned reasons and per-flag severity tell you which check fired. That is the project's stated position against the enterprise guardrail vendors it names by acquisition. Nothing here is a claim about how those thresholds hold against a determined attacker, so treat the returned probability as a signal to route on, with your own policy layered on top.
A Supabase project with the schema applied, which creates the api_keys, scans and demo_hits tables, plus TYPESAFE_API_KEY, NEXT_PUBLIC_SUPABASE_URL and SUPABASE_SERVICE_ROLE_KEY in your environment. Development runs on its own port via the Next binary. Authentication is Supabase magic link, and the SDK and CLI live in the separate jevegis-sdk repository rather than in this app.
Read them as a regression check. The eval command replays every labeled case against the live engine and reports verdict accuracy, per-flag misses, p50 and p95 latency and cost, currently 42 of 42, which is a small hand-labeled set and not a measure of adversarial robustness. Note also that usage logging is best-effort by design and never fails a scan, so a scan can succeed without leaving a log entry behind.
The /api/v1/scan and /api/v1/moderate routes require a bearer key and are logged; the unauthenticated /api/scan and /api/moderate routes are the playground behind the landing demo. Keys are issued self-serve at /api/keys and created or revoked at /api/dashboard/keys while signed in. Auth is Supabase magic link through /login, /auth/callback and /auth/finish, with session refresh in src/proxy.ts.
MIT licensed and open source, live at jevegis.vercel.app, with the SDK and CLI in a separate jevegis-sdk repository. It is a Next app where src/lib/engine.ts is the whole detection engine, request.ts parses text, direction or OpenAI-style messages, and apiAuth.ts does SHA-256 hashed bearer keys, a Postgres rate limit and best-effort usage logging.
Yes, by the project's own framing. The README argues that the 2025 acquisitions — Lakera to Check Point, Robust Intelligence to Cisco, Prompt Security to SentinelOne, Protect AI to Palo Alto — left standalone guardrails enterprise-only and demo-gated, and positions this as the self-serve, published-price alternative. So you call it over HTTP with a bearer key rather than vendoring the code.
More Jev resources like this
38GitHub stars
Python memory filter that judges every stored memory against the current request with one calibrated yes/no per memory in a single Jev request, keeping everything above a threshold.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
samdotmak/jev-recallTypeScript
samdotmak/jev-recallTypeScript0GitHub stars
Checks a text message for scam signals and turns them into an explainable verdict: code extracts link and sender signals, one parallel Jev call answers eleven typed questions, and a pure policy function weights them into a risk score.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
ShupingR/scam-shieldTypeScript
ShupingR/scam-shieldTypeScript30GitHub stars
Browser extension that filters your X timeline with a model you choose and pay for, using plain-language rules plus per-author and blocked-word controls.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
tshmieldev/sharpTypeScript
tshmieldev/sharpTypeScript23GitHub stars
Self-hosted Google Apps Script that classifies Gmail by meaning with Jev Choice over your own dozen-label taxonomy, checking metadata first and full content only when confidence demands it.
Jev Tools & CLIs#community#jevbest#catalog-import
36GitHub stars
Stock decision terminal that gathers Valyu market evidence across five categories and uses Jev for a typed 30-day outlook with probability, signals, and timing.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
unicodeveloper/jevocksTypeScript
unicodeveloper/jevocksTypeScript4GitHub stars
Playful chat demo that makes Jev hold a conversation one word at a time by picking from a compact 979-word vocabulary through repeated Choice calls.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
mkotlikov/jev-grugTypeScript
mkotlikov/jev-grugTypeScriptFrom the Jev guides
Original write-ups on what this entry is for.
The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.
Jev Tools & CLIs#gateway#vercel#cloudflare
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.
Jev Tools & CLIs#classification#email#security
Read article
From the Jev community
Posts from builders shipping with Jev right now.