jev-guardrails
TypeScript guardrail library plus DeepSeek Harness plugin screening prompts, tool calls, results, responses and claims with Jev.
Overview
dsh-jev-guardrails is a framework-agnostic TypeScript library that puts Jev in front of five points of an LLM application: prompts going into a model (jailbreaks, harmful requests, self-harm signals), tool calls before they run (destructive commands, exfiltration, remote code, credential access), tool results before they become context (prompt injection, hidden instructions, secrets), model responses before users see them, and claims checked against their evidence (verifyClaim). Jev returns probabilities and a severity score, and the library turns them into allow, review, block or support under a tunable policy.
The monorepo also includes a DeepSeek Harness bundle evaluating every tool call through tools/pre-execute, a client SDK and CLI with hook installers for OpenCode, Hermes Agent and DeepSeek Harness, and a Cloudflare Worker service offering paid API keys, prepaid credits and a SystemOne-compatible proxy. It works with the TypeSafe API, OpenRouter's Decisions API or the hosted service.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this prompt a jailbreak or harmful request? | — |
| Yes / no | Is this tool call destructive or exfiltrating data? | — |
| Yes / no | Does this tool result contain a prompt injection or secret? | — |
| Yes / no | Is this claim supported by its evidence? | — |
How it works
- Wrap prompts, tool calls, results and responses with the library
- Ask Jev typed questions at each point
- Policy maps probabilities and severity to allow, review, block or support
- Harness plugins apply the decision
Strengths
- Covers five screening points
- Multiple provider options
- Ready-made harness integrations
Caveats
- Hosted service is a paid third-party offering
- Policy tuning required
Getting started
`npm install && npm run build`, then `createGuardrails` from @codebam/jev-guardrails.At a glance
- Type
- Library
- Status
- beta
- Language
- TypeScript
- Runtime
- Library, DeepSeek Harness plugin, Cloudflare Worker
- Has code
- yes
Use cases and integrations
How curated lists describe it
Jev-backed guardrails for agent tool calls: library, native…
wh000wh000/awesome-jev-live
Frequently asked questions
What does Jev check?
Prompts, tool calls, tool results, responses and claims against evidence.
Who decides the action?
Your code, via a readable policy over Jev's probabilities.
Which harnesses are supported?
DeepSeek Harness natively, plus installers for OpenCode and Hermes Agent.
Related projects
Open-source agent guardrails that use Jev or LAYA to check prompts, retrieved content, tool calls and outputs for injection and risk.
ProjectbetaPythonPre-execution hook for coding agents that lets shell commands run, asks a human, or blocks them, with optional Jev classification.
ProjectbetaJavaScriptEffect-based safety gate for coding-agent shell commands: fast rules first, then Jev decides allow/deny plus eleven risk checks.
ProjectbetaTypeScriptDeepSeek Harness plugin where Jev classifies risk, irreversibility, task fit and injection for every tool call before it runs.
PluginbetaTypeScriptSecurity hook giving Codex, Copilot, Gemini, Cursor, pi, OpenCode and ACP an auto-mode-like guard via three Jev questions.
LibrarybetaTypeScriptSemantic MCP firewall that screens tool calls, results and descriptions with Jev, plus allowlists and session taint tracking.
MCP serverbetaTypeScript