Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Jev Sift

MCP tool that asks Jev which files, web pages or snippets are relevant to a query, so the agent reads selectively.

MCP serverbeta
View on GitHub ↗

Overview

Jev Sift is an MCP plugin that lets Jev decide what an agent should look at next. You pass a query and a batch of up to 50 items, each a file path, a public web page URL or a text snippet, and it loads the content, sends it straight to Jev, and returns compact relevance probabilities. The main agent then opens only the items worth a closer look, and full content from paths and URLs does not enter the agent's context first. Tool descriptions can be passed as text to judge whether calling a tool would help, without running it.

Instead of a single query you can send one to eight typed questions: boolean (mapped to Noul), choice, or score on an ordered rubric. Results keep input order and include per-item errors, truncation flags, resolved URLs and the model identifier. The README stresses that errors or truncation are not evidence an item is irrelevant. Setup needs only a JEV_API_KEY, or a private key file for desktop hosts; credentials never appear in tool responses and are not sent with web page fetches. A classify_status tool checks configuration without making a network call.

What the decision model decides

TypeQuestionAnswers
Yes / noDoes this item's content help with the stated task?yes · no
ChoiceUser-defined choice question per item—
ScoreUser-defined rubric score per item—

How it works

  1. Agent calls classify with a query and items
  2. Plugin loads files, pages or text
  3. Jev returns a relevance probability per item
  4. Agent opens only high or uncertain items

Strengths

  • Keeps bulky content out of agent context
  • Up to 50 items per call
  • Careful key handling

Caveats

  • Content is sent to the Jev API
  • Judges tool descriptions, not tool results
  • Truncation can hide relevant content

Getting started

Set JEV_API_KEY in the MCP server environment and call classify with a query and items

At a glance

Type
MCP server
Status
beta
Language
TypeScript
Runtime
MCP server
Has code
yes

Use cases and integrations

How curated lists describe it

  • MCP plugin that asks Jev which files, web pages, or text snippets are relevant to a query, so the agent reads selectively.

    MrJev/awesome-jev

Frequently asked questions

What does Jev decide?

How likely each file, page or snippet is to help with your query, or answers to your own typed questions.

Does the content enter my agent's context?

No. The tool loads and sends it to Jev; the agent sees only the results.

Should errors be treated as irrelevant?

No. The README says errors and truncation are not evidence of irrelevance.

Related projects

  • jev-filter

    npm CLI that screens large agent tool outputs with Jev, returning compact typed decisions and IDs while originals stay recoverable.

    CLIbetaTypeScript
  • alphaoptimizer

    Codex hook that compacts large command and tool outputs, using Jev to rank which chunks matter and keeping the full output retrievable.

    ServicebetaPython
  • ego-jev

    Agent skill where Ego Lite drives the browser and Jev resolves semantic choices: targets, filtering, ranking and evidence checks.

    ProjectbetaMarkdown + TypeScript
  • jev-carryforward

    MCP server with record and recall tools: facts are logged during a session and Jev keeps only those relevant to the next task.

    MCP serverbetaTypeScript
  • jev-rules

    Claude Code plugin where Jev asks one yes/no per standing rule and delivers only the rules relevant to each prompt, once per session.

    ProjectbetaJavaScript
  • jev-skill-gate

    Scores installed Claude Code skills against the current project with Jev and hides irrelevant ones via skillOverrides.

    Agent skillbetaJavaScript