User repository on GitHub
Keep skill catalogs outside the main LLM context. Jev selects relevant skills through one read-only MCP tool.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Agents & MCP
jev-skill-router (himomohi)
A local Python service that keeps your skill catalog out of the main model's context: it reads your SKILL.md library, asks Jev which skills fit, and returns only the selected instructions through a single read-only MCP tool, with Claude Code, Codex, and Cursor adapters.
- Stars
- 3
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- himomohi
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceRuns locallyRequires API key
Highlights
- Main-model footprint becomes approximately O(1) plus O(k loaded content) instead of O(N catalog metadata) plus O(k loaded content).
- Install with python Install.py --offline after Python 3.11+; it creates a private environment in ~/.jev-skill-router/runtime.
- uv path: uv tool install '.[secure]', then jev-skills setup --root ./examples/skills --offline and jev-skills route with a task.
- 0.4.1 adds MCP cancellation, explicit large-catalog retrieval, configuration-preserving upgrades, and tested-version release assets.
- Real Codex and Claude registration plus MCP connections passed; live model task quality remains unmeasured, per the README.
Quickstart
bash
uv tool install '.[secure]'
jev-skills setup --root ./examples/skills --offline
jev-skills route "Debug a Python traceback and failing tests"Watch out
Grounded in the repository README and GitHub metadata reviewed on 2026-10-15; the README's Codex setup section was truncated mid-sentence, no published PyPI package is assumed (pip install jev-skill-router is explicitly not the source-release instruction), and the project labels itself an early release.
Questions about this entry
Short answers drawn from the linked project or page.
Progressive-disclosure harnesses usually load skill names and descriptions at startup, and an index with hundreds of descriptions still consumes context. This project moves the selection inventory - not merely the bodies - outside the main model, so the main model sees a fixed router interface plus only the selected content.
A small local Python service reads your trusted SKILL.md library, asks TypeSafe Jev which skills actually fit the task, and returns only the selected instructions through one read-only MCP tool. Your existing agent still does the work with its normal tools and approval rules.
Download and extract the latest source ZIP, install Python 3.11 or newer, then run python Install.py --offline (Install.cmd on Windows, Install.command on macOS). The guided installer asks for a skill root and client, creates a private environment in ~/.jev-skill-router/runtime, and prints a stable CLI launcher.
No, and the README warns about this: adding an MCP server alone does not hide native skills. You must disable their native exposure or move user-managed skill folders out of the harness's discovery paths, and start a new session to avoid carrying the old inventory in conversation history.
No. Offline is an explicitly labeled keyword demonstration, not Jev: it validates the installation and loading path without pretending to reproduce model behavior. With no skill root supplied, the installer copies six bundled example skills into the app's data directory for that demo.
The README names Claude Code, Codex, and Cursor adapters, with English and Korean documentation. Real Codex and Claude registration and MCP connections passed; a practical Codex setup runs jev-skills setup --root ~/.agents/skills --client codex, then jev-skills auth and jev-skills doctor --live.
The project labels itself an early release: live Jev routing quality and end-to-end use in supported clients are still being validated. No published PyPI package is assumed - the README says pip install jev-skill-router is not the installation instruction - and hosted ChatGPT's built-in inventory cannot be removed by this local server.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Jev in front of the HeyGen MCP
X: Playwright CLI + Jev vs Playwright MCP
X: A local MCP server that cuts computer use from 23s to 0.8s. Cartwright built Cortex, a local MCP server where…
X: A side panel that drives any site's WebMCP tools
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
A Rust MCP server that recommends third-party agent skills: it retrieves deduplicated metadata with SQLite FTS5/BM25, sends one comparable typed Score per candidate in batches of at most 128 questions, and never installs or executes a skill.
Jev Agents & MCP#community#github#mcp
Community
king4s/jev-skill-routerPython
king4s/jev-skill-routerPythonClaude Code hook that asks Jev which installed skill fits each prompt and logs the answer, plus the script that joins that log to session transcripts and the negative result the author got from a week of running it.
Jev Agents & MCP#community#jevbest#catalog-import
Community
shimo4228/jev-skill-routerPython
shimo4228/jev-skill-routerPythonMore Jev resources like this
Lets AI agents use Windows applications through MCP or the command line: hand off a bounded task and Jev observes windows, chooses controls, types supplied text and navigates locally, returning a result or a request for help.
Jev Agents & MCP#community#catalog-import#logicrw
Community
jacks3tr/jev-desktopPython
jacks3tr/jev-desktopPythonA dependency-free MCP server that exposes TypeSafe's Jev to Claude Code, Codex, Cursor, Hermes, and other agents through the ask_jev and list_jev_models tools, so your code gets typed decisions with probabilities instead of prose to parse.
Jev Agents & MCP#community#catalog-import#logicrw
Community
codeia-academy/jev-mcpPython
codeia-academy/jev-mcpPythonA local stdio MCP server that adds a jev_decide tool to OpenCode agents, supporting noul, choice and score questions so an agent can route or gate work on structured answers and probabilities instead of prose.
Jev Agents & MCP#community#catalog-import#logicrw
Code Mode for MCP where the sub-model is a calibrated decision model instead of an LLM: agents get search(query) and execute(code), and Jev makes per-item choices in about 100 ms, benchmarked on MCPToolBench++, LiveMCPBench and When2Call.
Jev Agents & MCP#community#catalog-import#logicrw
Community
vishiatchoudhary/tooljevPython
vishiatchoudhary/tooljevPythonAn MCP relay that adds use_jev to every configured server: Jev picks the tool calls toward a goal, the calling model writes any text values, and the relay executes the steps and returns a trace.
Jev Agents & MCP#community#catalog-import#logicrw
Community
chy4pro/jev-in-mcpTypeScript
chy4pro/jev-in-mcpTypeScriptAn open-source Jev-based MCP router: it discovers, catalogs, classifies and deduplicates every MCP server and tool you have, then exposes each agent only the small, authorized subset it needs for the task. Routing decisions are System One Choice/Score/Noul questions, scored by a pluggable engine at the edge or hosted.
Jev Agents & MCP#community#catalog-import#logicrw
Community
amantux/mcp-routerPython
amantux/mcp-routerPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.