kerpopule/hermes-jev-skills

What it solves

AI agents often waste expensive frontier-model tokens on simple decision-making tasks—such as selecting the right model for a turn, filtering search results, or sorting mail—that do not require complex prose generation. This project provides a set of skills and a plugin to offload these "small decisions" to Jev, a specialized, low-cost decision model that answers typed questions (yes/no, scoring, or picking from a list) with calibrated confidence, reducing latency and cost.

How it works

The system integrates Jev into an agent's workflow via a plugin (specifically for Hermes) or standalone skill files. It intercepts agent processes at key points to make rapid decisions:

• Model Routing: Dynamically selects the most cost-effective model capable of handling the current turn.
• Filtering & Selection: Identifies which retrieved memory passages or search results are actually relevant before the main LLM reads them.
• Task Triage: Categorizes incoming messages by urgency and type.
• Action Selection: Picks the next GUI or browser action from a pre-defined list of safe actions.

To ensure safety and privacy, the system uses a redaction layer to mask sensitive data before sending it to the Jev API and implements a "fail-open" design where the agent reverts to default behavior if the decision model is unavailable.

Who it’s for

Developers building AI agents (particularly those using the Hermes agent framework) who want to optimize their operational costs and response times by separating high-level reasoning from routine decision-making.

Highlights

• Cost & Speed: Decisions typically take ~0.4s and cost a fraction of a cent.
• Shadow Mode: Allows users to log Jev's decisions without actually switching models or actions to verify performance before going live.
• Privacy-First: Includes built-in redaction for emails, phones, and tokens, with a "private profiles" mode for highly sensitive data.
• Broad Utility: Ships with nine agent-agnostic skills that can be used in other environments like Claude Code or Codex.
• Fail-Safe: Designed to fail open, ensuring that Jev outages never block an agent's turn.