jev-scout

Zero-hallucination repo & crate discovery. We take soundings, not guesses: real registry depth, typed scores, verified install routes.

$ cargo install jev-scout read the chart ↓

DATUM

FIG A sounding: "fast sqlite tui in rust" SCOUT_REPOS

verified sounding filtered (fit < 2.5)

One fan-out call, five candidates, typed verdicts. Absence is as informative as presence: weak matches stay ghosted, never deleted.

LLMs invent packages.
We take soundings.

Ask a general model for a library and it will happily recommend a name that never shipped. GitHub keyword search can't parse a concept. A hydrographic chart solves both: every plotted point is a real registry measurement, every hazard is marked.

SB-01

Grounding first

Two real registries, one pass: GitHub REST + crates.io. Live metadata only: stars, downloads, pushed_at, language, topics. Nothing imagined.

SB-02

One fan-out call

All candidates scored in a single speculative fan-out to TypeSafe Jev System One. Score (fit 1-4), Noul (maintenance probability), Choice (best match). One request, typed answers.

SB-03

Deterministic ranking

score × confidence, sorted. Recency decay ×0.9 past 180 days. Weak matches filtered at fit < 2.5 or confidence < 0.5: visible, not silent.

SB-04

Dual surface

Terminal CLI for you, stdio MCP for your agent. The same soundings, one JSON-RPC 2.0 contract, structuredContent out.

Survey instruments

Ten flags, charted. Every one returns measurable depth, never an essay.

jev-scout "fast sqlite tui in rust"conceptual query → ranked verified candidates

-e / --ecosystem all|github|cratesrestrict the survey water

-n / --limit 5plot depth max 10

-j / --jsonmachine-readable soundings for scripts and agents

--no-filtershow ghosts too (no one shoots the messenger)

--mcpstdio MCP server · scout_repos tool · structuredContent

--help / --versionthe chart legend and datum

The oracle is a sextant, not a novelist

Jev answers machine-native questions with machine-native types. It will not write you a confident paragraph you then have to verify.

Scorearchitectural fit, 1-4: does this repo actually do the thing

Noulcalibrated probability of active maintenance, from pushed_at and topics

Choicebest_match: the single candidate the survey would send you to

Fan-outone state, many questions, one request. ~10x fewer round trips

Jev refuses arithmetic and word counts, and it is right to: the CLI does the math, Jev does the judgment. That split is the whole architecture.

Notices to Mariners

Corrections are how a chart earns trust. These are the honest numbers, measured, not marketed.

LATENCY: COLD

~2.0-2.3s

Two serial external legs: GitHub API ~0.85s + Jev ~1.0s floor. No 1.5s marketing figure here.

LATENCY: WARM

~0.3s

Same process, TTL 60s caches on search and evaluation. This is the agent-relevant number.

RECENCY DECAY

×0.9 / 180d

Stale candidates lose weight. Julian-day math in host code: Jev never counts.

FILTER THRESHOLD

fit 2.5 / conf 0.5

Below the line, ghosted on the chart. --no-filter shows the ghosts.

Set up in two lines

# 1: install
$ cargo install jev-scout

# 2: free oracle key
$ export TYPESAFE_API_KEY=your_key_here

# 3: agent surface (Claude Code)
$ claude mcp add jev-scout -- bash -c "export TYPESAFE_API_KEY=$TYPESAFE_API_KEY && jev-scout --mcp"

One static binary under 5MB. No SQLite, no Chromium, no browser automation, no paid search proxy, no text generation. Verified soundings only.

enter the harbor ↗ crates.io