jkudish/jev-mcp
Fast, cheap, typed judgments from TypeSafe's Jev model, as MCP tools.
What it solves
Agents often skip "mechanical" checks—like verifying facts, screening content, or sorting lists—because frontier models are too slow and expensive to run on every small task. This project provides a set of specialized, fast, and low-cost tools that allow agents to perform these judgments reliably and at scale.
How it works
It acts as an MCP (Model Context Protocol) server that connects your agent to TypeSafe's Jev model. You register these tools with your agent, which can then call them to perform specific tasks like fact-checking, content screening, or data extraction. Each tool returns structured, typed data, including confidence scores and probabilities, allowing the agent to decide whether to trust the result or flag it for human review.
Who it’s for
This is for developers building AI agents who need to add robust, verifiable decision-making capabilities to their workflows without the latency or cost of using a full-scale frontier model for every minor step.
Highlights
- Twelve specialized tools: Includes functions for verifying claims (
jev_verify), screening for prompt injections (jev_screen), ranking candidates without embeddings (jev_find), and batch-classifying items (jev_classify). - Fast and cheap: Judgments are returned in 150 to 500 milliseconds for a fraction of a cent.
- Typed outputs: Provides clear confidence scores and probability distributions, enabling agents to handle uncertainty gracefully.
- Flexible deployment: Runs locally via stdio, as a remote HTTP server for teams, or embedded directly into your own application code.
- Agent-ready: Includes a pre-built agent skill that teaches coding assistants when to use these tools instead of relying on their own internal knowledge.
Related
- Dispatch
- Project
- Project
- Project
- Project