jkudish/jev-mcp

Fast, cheap, typed judgments from TypeSafe's Jev model, as MCP tools.

What it solves

Jev MCP provides a suite of fast, low-cost, and typed judgment tools for AI agents. It addresses the problem where frontier models are too slow or expensive to perform routine mechanical checks—such as fact-checking, content screening, and reranking—on every piece of data an agent encounters.

How it works

It acts as a Model Context Protocol (MCP) server that connects agents to TypeSafe's Jev model. The server exposes twelve specialized tools that return structured, typed data (like probabilities and confidence scores) rather than free-form text. These tools operate in the 150-500ms range and are designed to be used as "guardrails" or filters before data enters an agent's main context.

Who it’s for

Developers building AI agents who need to implement high-frequency verification, classification, or routing tasks without the latency and cost of using a large frontier LLM for every small decision.

Highlights

  • Twelve Specialized Tools: Includes tools for claim verification (jev_verify), prompt injection screening (jev_screen), calibrated probability checks (jev_noul), and semantic finding (jev_find).
  • Typed Judgments: Returns concrete probabilities and confidence scores, allowing agents to programmatically decide whether to trust a result or flag it for human review.
  • High Performance: Designed for speed (150-500ms) and low cost (fraction of a cent per call).
  • Flexible Deployment: Supports standard stdio transport, stateless HTTP mode for remote teams, and in-process embedding via a Node.js package.
  • Agent Skills: Includes pre-defined skills to teach coding agents exactly when to use these judgment tools instead of relying on their own internal reasoning.

Related

  • Dispatch
  • Dispatch
  • Project
  • Dispatch
  • Project