How accurate is the skill selection actually?

On the author's harness with about 130 installed skills: it picked the right installed skill for 9 of 10 session goals, said so when the right skill only existed on skills.sh in 4 of 5 cases, stayed quiet when no skill would help in 3 of 4, and went 10 of 10 on single prompts. None of the test prompts name the expected skill, and the misses are kept and explained in the report.

Why not just ask Claude which skill to use?

Because that means writing a prompt, parsing the text back and trusting whatever confidence the model chose to write. Jev takes a list of options and returns a calibrated probability for each in a single evaluation, cannot produce a malformed answer, reports when it is unsure, and costs about two orders of magnitude less than a chat model call.

Will it fire on every prompt and get in the way?

Not by default. The first version did run on every prompt and was accurate on tests but irritating in practice, because once a session was about skills it kept suggesting the skill-authoring skill. The default is one decision at session start, per-prompt mode is an opt-in through install-hook with the UserPromptSubmit event, and you retarget mid-session with the intent command.

What does it send where, and will it install skills on its own?

It installs nothing: it prints install commands and tells Claude to ask you first, and when nothing installed fits it hands off to the find-skills search without suggesting candidates, so that skill's own quality filter decides. There is no server and no telemetry of its own, but your goal text and directory name go to TypeSafe's API and a search string goes to skills.sh through npx skills.

What do I need to install it?

Python 3.12 and uv, then clone the repo, run uv sync, and copy .env.example to .env with your TypeSafe key; register the session hook per project or in your user-level Claude settings.

Can I try it without a TypeSafe key?

Yes. Without an API key it runs in mock mode, matching keywords instead of asking Jev and labelling its output mock; the tests use that mode too, so no key is needed to try the tool.

Where does it look for installed skills?

Project .claude/skills, user ~/.claude/skills, enabled plugins under ~/.claude/plugins/cache, skills bundled with the Claude desktop app, and built-in skills compiled into Claude Code; switched-off skills are excluded.

CommunityJev Agents & MCP

Skill Router

Claude Code CLI that ranks every skill installed on your machine against a session goal using Jev, then tells the agent which one to load. Three possible answers: load this skill, nothing you have installed fits, or no skill is needed, and nothing else runs for the rest of the session.

See highlights

Open github.com

Stars: 4
Language: Python
License: MIT
Published by: Community
Author: lomeshdutta
Last verified: Oct 2026

lomeshdutta/skill-router

User repository on GitHub

View on GitHub

Tell Claude Code which installed skill a session needs, using Jev (TypeSafe AI) for the decision and skills.sh for discovery.

Forks: 2
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A small command-line tool for Claude Code that picks the right skill for your session. Claude Code lets you install skills: folders of instructions that teach it a specific job, such as writing cold emails, auditing a page for SEO, reviewing a diff, or building a PowerPoint deck. They come from the skills.sh directory, from plugins, and from Claude Code itself. Once you have more than a dozen, three things go wrong: You forget which ones you have, so you do the work without the skill that would have helped.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Jev returns a probability per installed skill plus a confidence score, typically well under a second, for a fraction of a cent.
On a machine with about 130 installed skills, evals picked the right installed skill for 9 of 10 session goals and 10 of 10 single prompts.
It scans project, user, plugin cache and macOS desktop app skill locations, and excludes skills you have switched off in settings.
It matches meaning rather than words: trigger vocabulary about an unrelated topic scored 0.00, a paraphrase sharing no words scored 0.92.

Watch out

The routing accuracy figures come from the author's own eval harness on one machine with roughly 130 installed skills; reproducing them needs a TypeSafe key and a similarly sized skill set, and the mock mode the tests use scores nothing meaningful by design. Latency varies sharply by location: about 0.35 s from the US West Coast and about 5 s on an independent run elsewhere, which matters if you expect this on every session start. Coverage is Claude Code's only, since the hook contract and skill locations are its own, and your goal text and directory name go to TypeSafe's API while search strings go to skills.sh. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-recruiter

Local LinkedIn recruiting workspace on Browser Use's Jev Ultrafast: Jev screens titles, drives navigation, and assesses profile excerpts against your brief while a live browser stays visible; matches and evidence export as JSON.