What does toolJev do?

A Code Mode layer for MCP where the sub-model is a decision model rather than an LLM. It sits in front of your MCP servers and gives the agent two tools, search and execute, instead of every server schema.

How does Jev make per-item decisions in code?

Inside execute(code), the agent's Python calls tools as mcp. . () and makes judgments with jev.choice, noul or map. Those return typed answers in about 100 ms without an LLM turn, which is Code Mode crossed with a recursive language model.

What were the headline benchmark numbers?

Jev ranked the right tool first on 83% of MCPToolBench++, 54% of LiveMCPBench and 99% of When2Call; retrieval alone scored 73%, 40% and 92%, and Jev with no retrieval just 18%.

How does it help agents with hundreds of tools?

With 612 tools the agent called the right tool in 89% of tasks versus 81% for Claude Code's own tool search, used 77% fewer input tokens and cost 28% less, though it was slower: 34 s versus 18 s.

What about routing support tickets?

On 400 tickets Jev alone was 98.8% accurate, and on the 97% of answers where it was at least 0.9 confident it was 99.7% accurate. An agent routing with Jev cost $0.063 against $0.332 with Haiku doing every ticket, 5.3x cheaper at about the same accuracy.

When is direct tool calling better?

Below roughly 100 tools. With 87 tools, direct tool calling was more accurate (0.94 versus 0.89) and cheaper because it used a local backend, so the README advises not bothering with the layer at that scale.

What are the caveats?

The author lists them in bench/RESULTS.md: routing numbers use up to 200 answerable and 200 unanswerable queries per benchmark, agent runs use 1-2 reps and 36 tasks, upstream servers are mocks built from tool schemas, and the 87-tool run used the much less accurate nanojev.

Jev directory

CommunityJev Agents & MCP

toolJev

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.

See highlights Open github.com

Stars: 3
Language: Python
License: MIT
Reading time: 3 min
Published by: Community
Author: vishiatchoudhary
Last verified: Oct 2026

vishiatchoudhary/tooljev

User repository on GitHub

View on GitHub

Code Mode for MCP, where the sub-model is a calibrated decision model (Jev), not an LLM. Benchmarked on MCPToolBench++, LiveMCPBench, When2Call and live Claude agents.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity catalog-import logicrw mcp code-mode benchmark

Works forJev for Tool calling Jev for Routing Jev for Evaluation

Used inJev for Software & DevTools Jev for Customer Support

Languages: Python
Platforms: MCP
Capabilities: Open sourceRequires API key

Highlights

Two tools replace hundreds: search(query) shortlists MCP tools and Jev says whether any fits; execute(code) runs agent Python calling mcp.<server>.<tool>().
With 612 tools the agent called the right tool in 89% of tasks versus 81% for Claude Code's own tool search, with 77% fewer input tokens but slower.
Routing 400 support tickets cost $0.063 at 98.8% accuracy, against $0.332 at 99.3% when Claude Haiku routed each ticket itself.
Jev ranked the right tool first on 83% of MCPToolBench++, 54% of LiveMCPBench and 99% of When2Call; retrieval alone scored 73%, 40% and 92%.
Below about 100 tools direct calling won: with 87 tools it was more accurate (0.94 versus 0.89) and cheaper with a local backend.

Quickstart

bash

uv run python examples/showcase.py

Watch out

Grounded in the truncated repository README and GitHub metadata reviewed on 2026-10-15; the README was cut off mid-figure. The author's own caveats note 1-2 reps on 36 tasks for agent runs, mock upstream servers built from benchmark tool schemas, and that the 87-tool run used nanojev, a much less accurate local stand-in.

Questions about this entry

Short answers drawn from the linked project or page.

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

From the Jev guides

Original write-ups on what this entry is for.

MCP tool routing: Jev in front of the tool call

Oct 16, 2026

Tool menus can outgrow the context that reads them. The catalog's Jev-based routers and gates narrow the list: benchmarked tool search, per-tool scoring, fail-closed approvals, and WebMCP in the browser.

Jev Agents & MCP#mcp#tool-routing#agents

Read article

Approval policies for coding agents: hooks, guards, and gateways

Oct 16, 2026

How the Jev guard family turns coding-agent permissions into auto-approve, deny, or ask: hooks that enforce, guards that fail safe, and gateways that gate generation.

Jev Agents & MCP#coding-agents#guardrails#hooks