JevCore

•
110
•

Empower AI agents with TypeSafe Jev, delivering precise, typed judgments and calibrated probabilities for critical decisions, operating offline by default with transparent egress controls.

JevCore provides TypeSafe Jev, a specialized decision model designed for AI agents, DeepSeek Harness, and other Model Context Protocol hosts. Unlike general chat models, Jev focuses on answering precise, typed questions—such as yes/no, multiple choice, or scoring—and returns calibrated probabilities to inform agent decisions. This project prioritizes security and privacy by operating offline by default, requiring explicit opt-in for any network communication, clearly disclosing all data egress, and preventing models from altering their own constraints. JevCore is available as a core library, a DeepSeek Harness plugin, and an MCP server, ensuring consistent security and reliable decision-making across various deployment environments.

Características Principales

010 GitHub stars
02Offline by default with explicit egress disclosure for network calls.
03Provides typed judgments (yes/no, choice, score) with calibrated probabilities for AI agent decisions.
04Offers robust security guarantees, preventing model reconfiguration of gates and ensuring sensitive data redaction.
05Integrated as a DeepSeek Harness plugin, a standalone Model Context Protocol (MCP) server, or a core library.
06Configurable safety and context gates to control tool execution and information flow.

Casos de Uso

01Integrating TypeSafe Jev decision-making into DeepSeek Harness workflows.
02Serving Jev tools via a standard I/O-based Model Context Protocol server for any compatible host.
03Implementing robust decision layers and security gates for AI agents in plain Node.js applications.