Jev Scout — AkashPriyadarshii

Discovery and decision are separated. Grounding queries the GitHub REST API and crates.io for live stars, licenses and commit dates, then one speculative System One call scores every candidate with Score for architectural fit, Noul for active maintenance and Choice for the best match. Ranking is deterministic: score times confidence.

Names come from the registries, not from a model, so a hallucinated package name is not the failure mode here. The judgement call is ranking: fit is scored 1 to 4, maintenance is a probability, and weak matches are filtered out unless you pass --no-filter to see every candidate. It also writes no summaries or essays.

A cold process takes about 2.0-2.3s, and the floor is two serial external legs: GitHub search around 0.9s plus a 1.0s Jev call, the raw curl floor even for one question. Warm, inside an MCP session with 60s in-process caches, repeats come back in roughly 0.3s. Results are capped at 1 to 10 per call.

It runs as a stdio JSON-RPC 2.0 MCP server: point a client at jev-scout --mcp with TYPESAFE_API_KEY exported. The scout_repos tool takes query plus optional ecosystem (all, github, crates), limit from 1 to 10, and strict filtering of weak matches, on by default. Offline unit tests run against recorded fixtures.

Install Rust and the CLI with `cargo install jev-scout`, export TYPESAFE_API_KEY, then run `jev-scout "fast sqlite tui in rust" --json`. It is MIT-licensed, and every live search calls the TypeSafe endpoint plus live GitHub and crates.io metadata.

It ranks by confidence-weighted score, Score for architectural fit times confidence, and returns only verified repositories with exact stars, licenses and clone commands. It writes no summaries or essays, and fit is scored on a 1 to 4 scale.

The scout_repos MCP tool takes query plus ecosystem (all, github or crates), a limit from 1 to 10, and strict filtering of weak matches, which is on by default. Passing --no-filter shows every candidate instead.