jev-skill-router
Local Python MCP service that keeps the skill catalog out of the agent's context and lets Jev pick which SKILL.md files to load.
Overview
Jev Skill Router is a small local Python service that reads a trusted SKILL.md library, asks TypeSafe Jev which skills fit the current request, and returns only the selected instructions through a single read-only MCP tool. The agent keeps doing the work with its normal tools and approval rules.
It targets users of Claude Code, Codex and Cursor whose skill catalogs have grown large enough that even names and descriptions consume context. The design moves the selection inventory out of the main model so its footprint is a fixed router interface plus the selected content. It ships a guided installer, keychain-only key storage, a `park` command that moves native skill folders to a vault, and an offline keyword mode that is explicitly labelled as not Jev. The README calls it an early release whose live routing quality is still being validated.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which skills in the catalog fit the user's request?(inferred) | — |
How it works
- Agent calls the router's single MCP tool with the task
- Service reads SKILL.md metadata from the configured roots
- Jev judges which skills fit
- Only the selected skill instructions are returned to the agent; offline mode uses keywords instead
Strengths
- Main-model context cost becomes roughly constant instead of growing with catalog size
- API key kept only in an OS keychain or server environment, never a plaintext file
- Adapters for Claude Code, Codex and Cursor plus English and Korean docs
Caveats
- Early release; live Jev routing quality and end-to-end client use are still being validated
- Native skills must be disabled or parked separately, or the agent still sees them
- No published PyPI package; install is from source
Getting started
Download the source ZIP and run `python Install.py --offline`, or with uv: `uv tool install '.[secure]'` then `jev-skills setup --root ./examples/skills --offline`.At a glance
- Type
- MCP server
- Status
- beta
- Language
- Python
- Runtime
- MCP server
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Keep skill catalogs outside the main LLM context. Jev selects relevant skills through one read-only MCP tool.
logicrw/awesome-jev-projects
Frequently asked questions
How is this different from progressive skill disclosure?
Harnesses still load every skill's name and description; this router moves that whole inventory to a separate service so the main model sees only the selection.
Does the offline mode use Jev?
No. Offline mode is a labelled keyword demonstration for checking installation, not a reproduction of Jev's behaviour.
Do I need a TypeSafe key?
Yes for live routing; it is stored via `jev-skills auth` in the OS keychain or supplied as TYPESAFE_API_KEY to the server.
Related projects
Toolkit where Jev condenses large MCP responses for Claude Code and maps Codex requests to reviewed, fixed repository operations.
ProjectbetaTypeScriptGives an agent one tool that uses Jev to search a nested capability tree and return only the documented commands a task needs.
ProjectprototypeTypeScriptMCP server exposing Jev Noul, Choice and Score via OpenRouter so coding agents can make quick decisions without reasoning tokens.
MCP serverbetaTypeScriptMCP relay that adds use_jev to every server: Jev picks tool calls, the calling model writes free text, the relay executes.
MCP serverbetaTypeScriptOne-command npx installer for a set of Jev-powered agent tools: context compaction, browser agent, prompt enhancer and skill router.
ProjectbetaJavaScriptLocal gateway, MCP server and Claude Code hooks that let coding agents delegate explicit choices, classification and scoring to Jev.
MCP serverbetaTypeScript