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

jev-scout

MCP server that scores every search query, result and fetched page an agent sees for relevance, credibility and steering risk.

MCP serverbeta
View on GitHub ↗

Overview

jev-scout is an MCP stdio server exposing guarded_search, guarded_fetch and research_report, in which every query, result and fetched page is scored by Jev with atomic, observable-signal criteria, and every decision streams to a live local dashboard. The motivation is that agents research the web through opaque tools, so nobody sees which sources they trusted or what steered them; Jev's low per-decision cost makes it viable to judge every step.

Each result gets one Jev call for relevance, credibility and whether it is worth fetching; each fetched page gets one for whether it answered the question, its content class and steering risk. Decisions go to a JSONL ledger and a dashboard with a live feed (with an explicit unscored badge when Jev is unreachable), a cost ticker and a per-domain credibility table. It is deliberately not a research agent: no planning, synthesis or answer writing; it annotates or optionally gates what an agent sees. Rubrics were validated on a hand-labelled 25-item golden set at 88% relevance and 96% credibility, with misses analysed. It uses SearXNG by default and includes session budgets and SSRF-guarded fetching.

What the decision model decides

TypeQuestionAnswers
Yes / noIs this search result relevant / credible / worth fetching?yes · no
Yes / noDoes this fetched page answer the question?yes · no
ChoiceWhat class of content is this page?—
Yes / noDoes this page try to steer the agent?yes · no

How it works

  1. The agent searches through guarded_search
  2. Jev scores each result; optionally gate low scores
  3. guarded_fetch retrieves pages with SSRF protection
  4. Jev scores each page for answer, class and steering risk
  5. Decisions stream to the ledger and dashboard

Strengths

  • Makes agent research visible
  • Explicit unscored status instead of misleading zeros
  • Validated rubrics with miss analysis
  • Session budgets and SSRF guards

Caveats

  • Small 25-item golden set
  • Requires SearXNG setup and a Jev key
  • Depends on a sibling jev-client library

Getting started

Start SearXNG with JSON enabled, clone jev-client and jev-scout side by side, build, then claude mcp add --scope project --transport stdio jev-scout -- node --experimental-strip-types /ABS/PATH/jev-scout/src/server-mcp.ts

At a glance

Type
MCP server
Status
beta
Language
TypeScript
Runtime
Node MCP server with local dashboard
Has code
yes

Use cases and integrations

How curated lists describe it

  • MCP server that scores an agent's every search query, result, and fetched page for relevance and credibility, with session budgets, SSRF-guarded fetching, and a live decision dashboard.

    cobanov/awesome-jev

  • Hand-labeled 25-item search-triage study with pinned rubric versions and a drift baseline; reports 88% relevance and 96% credibility with all four misses decomposed.

    cobanov/awesome-jev

Frequently asked questions

What does Jev decide?

Relevance, credibility and worth-fetching for results, and answer, class and steering risk for pages.

Is it a research agent?

No. It is a decision layer in front of search that annotates or gates results.

How accurate are the rubrics?

88% relevance and 96% credibility on a 25-item hand-labelled set.

Related projects

  • ReadyCode Reader

    Local-first MCP server that returns only Jev-checked passages from big PDFs, Word files and spreadsheets, with exact row calculations.

    MCP serverbetaTypeScript
  • jev-research-mcp

    Self-hosted MCP research server that filters search results and page blocks with Jev, returning compact evidence with provenance.

    MCP serverbeta
  • jev-seo

    Experimental Rust SEO and GEO CLI and MCP server combining page checks, DuckDuckGo queries and optional Jev scoring.

    MCP serverprototypeRust
  • jev-playwright-mcp

    Drop-in proxy for Playwright MCP that adds Jev page triage, prompt-injection masking, snapshot pruning and risky-action gating.

    MCP serverbetaTypeScript
  • jev-security-scan

    Reviews agent skills, MCP configs and source with local static checks plus Jev before install, reporting evidence, risk and coverage gaps.

    MCP serverbetaPython
  • jev-shield

    Semantic MCP firewall that screens tool calls, results and descriptions with Jev, plus allowlists and session taint tracking.

    MCP serverbetaTypeScript