Overview & result
An MCP server that gives coding agents a decision router backed by Jev. It evaluates context, code diffs, logs or planning options and returns calibrated probabilities in three formats: noul (binary), choice (categorical) and score (rubric). The author reports about 150 ms per decision through the OpenRouter Decisions API and zero output tokens per call, and lists Cursor, Windsurf, Claude Desktop and Claude Code as clients.
From the original post
open-sourced jev-flash-router: an MCP server for TypeSafe's new Jev model. AI coding agents waste hundreds of reasoning tokens just deciding which file to edit, which route to pick, or whether a diff breaks tests. Jev evaluates state and outputs calibrated probabilities. Works with Cursor, Windsurf, & Claude Code
How Jev fits in the loop
- The agent calls the evaluate_decision tool with a state (context, diff, log or task description) and a targeted question.
- It chooses a format, noul, choice or score, and supplies the criteria.
- The server sends the request to Jev through the OpenRouter Decisions API.
- Calibrated probabilities come back with no output tokens.
How to reproduce
- Inspect the original showcase and source material at https://github.com/Ravinder82/jev-flash-router
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It exposes Jev to coding agents as a single tool, so questions like "will this change break the API?" cost input tokens only.
Limitations
- It needs an OpenRouter API key with access to the Jev model.
- Latency and cost figures (about 150 ms, about $0.042 per million input tokens) are the author's.
- The README gives no accuracy evaluation.