jev-judge-mcp
An MCP server that gives a coding agent eleven typed judgment tools backed by TypeSafe's Jev model, answering fixed-answer questions such as whether a claim is supported or a page is safe to read, then routing the result to auto-accept, review or escalate.
jev-judge-mcp is an MCP server that gives a coding agent judgment tools backed by TypeSafe's Jev model, for steps that have a fixed set of possible answers rather than needing new text or code. The agent hands a tool some evidence and a question it can enumerate, such as whether a claim is supported, whether a page is safe to read, which of several files answers a question, or whether a patch finished a task, and Jev answers with a probability.
A policy layer turns that probability into one of three actions: proceed automatically, check the result another way, or stop and escalate to a human or the calling LLM. The project reports a median round-trip of under half a second and a cost of about $0.025 per thousand decisions on its recorded benchmark, with every auto-accepted answer on that run correct and misses routed to review instead of passing silently.
It requires Python 3.12 or newer, uv, a POSIX system, and a TypeSafe API key, and is installed from PyPI under the MIT license.