GitHub · by Joey Kudish

jev-mcp

An MCP server with claim checks, screening and ranking built on Jev.

Joey Kudish’s proof-of-concept MCP server for Jev. It wraps fact checking, prompt-injection screening and semantic ranking as tools an agent can call.

There are ten tools rather than one general-purpose call, and the list is the argument: `jev_verify` checks claims against evidence, `jev_screen` judges content before it enters context, `jev_find` picks the best candidate by meaning, `jev_rerank` scores and sorts every candidate, `jev_classify` batch-assigns items to classes, `jev_decide` settles bounded alternatives, `jev_compare` judges how two passages relate, `jev_extract` pulls field values with regex plus judgment, `jev_review` scores a proposed diff, and `jev_gate` reviews a patch and verifies completion claims in one call.

Each judgment comes back typed — probabilities, and for most tools a confidence score — in roughly 150 to 500 milliseconds for a fraction of a cent. Joey Kudish frames those as the cheap mechanical checks agents otherwise skip, because a frontier model is too slow to run on every page, every claim and every candidate list.

Two of the uses are worth pulling out because they are hard to get any other way. Screening a fetched page for injected instructions before it reaches context is a defence that has to run on every page or it is not a defence. And finding which of several hundred documents answers a question, with no embeddings and no index to maintain, removes an entire piece of infrastructure from a retrieval stack. It needs Node 20 or newer, and the author calls it early software.