jevegis
Hosted guardrails API checking LLM inputs, outputs and retrieved documents for injection, jailbreaks and leaks in one Jev call.
Overview
Jevegis provides guardrails for LLM apps in a single API call. It covers security checks for prompt injection, jailbreaks, credential and PII leaks, unauthorised actions, malicious code and indirect injection in retrieved documents, plus trust-and-safety moderation, all powered by Jev's typed judgments with calibrated probabilities and sub-second calls. The whole detection engine is one declarative list of checks, each phrased appropriately for its target (input, output, document or content); an evaluate function builds one Jev request, applies thresholds and returns a verdict with reasons, flags, category and severity.
It accepts plain text, a direction, or OpenAI-style messages where the last message is judged and the rest is context, with custom thresholds and check selection. Authenticated scan and moderate endpoints use hashed bearer keys, Postgres rate limiting and best-effort usage logging, and an unauthenticated playground powers the landing demo. Supabase handles auth with magic links and a dashboard for key management. An eval script runs labelled cases against the live engine and reports accuracy, per-flag misses, latency and cost; the README reports 42 of 42. It is live on Vercel, MIT licensed, with a separate SDK and CLI.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this input contain a prompt injection or jailbreak attempt? | — |
| Yes / no | Does this output leak credentials or PII? | — |
| Yes / no | Does this retrieved document contain indirect injection? | — |
How it works
- Client posts text or messages to /api/v1/scan or /moderate
- Engine builds one Jev request from the selected checks
- Thresholds produce a verdict with flags, category and severity
- Usage logged; dashboard manages keys
Strengths
- Many guardrail checks in one call
- Configurable thresholds and checks
- Eval suite reporting 42/42 per the README
- Open source with hosted option
Caveats
- Eval set is small (42 cases)
- Hosted service depends on Supabase and TypeSafe
Getting started
Copy .env.example to .env.local (TYPESAFE_API_KEY, Supabase settings), then run the Next dev server.At a glance
- Type
- Service
- Status
- beta
- Language
- TypeScript
- Runtime
- Next.js API (Vercel) + Supabase
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Guardrails for LLM apps in one API call. Prompt injection, jailbreaks, leaks, unsafe content. Built on TypeSafe Jev. MIT.
hellogumbo/awesome-jev
Frequently asked questions
What does Jevegis check?
Prompt injection, jailbreaks, credential and PII leaks, unauthorised actions, malicious code, indirect injection and unsafe content.
How many API calls per check?
One Jev request covers all selected checks.
How is it evaluated?
An eval script runs labelled cases; the README reports 42 of 42 correct.
Related projects
Example of using Jev as a webhook guardrail inside the agentgateway proxy to inspect model requests and responses.
ProxyproductionRustGateway admission control that asks Jev what each incoming request is trying to do, filtering prompt injection and abuse.
AppbetaLuaGitHub Action that asks Jev yes/no questions about each diff chunk for secrets, vulnerabilities, complexity and prompt injection.
ProjectbetaPythonJev Sentinel: guard for Pi, Claude Code and Codex that checks tool calls, tool outputs and replies, failing closed.
PluginbetaTypeScriptCLI gate for X drafts: local hard secret checks, then Jev Noul questions with thresholds; exits non-zero to block posting.
ProjectbetaPythonGo reverse proxy between nginx and an LLM backend that extracts each request's user input, asks Jev if it is harmful, and blocks or passes.
ProxybetaGo