How is this different from normal LLM tool calling?

The writer has no tool bindings, catalog, schemas, or routing choices and cannot select, approve, or stop the loop. Only Jev selects actions, Python validates arguments, and rejected arguments are never executed.

Does the token accounting prove it saves tokens?

No. The repo says the design reduces sources of LLM token use but does not establish a measured saving against a conventional agent. Per-slot generation and Jev argument approval add latency and API calls.

Which keys does a run need, and must the writer stay on gpt-4.1-mini?

Two keys in .env: TYPESAFE_API_KEY for Jev, which defaults to jev-latest, and OPENAI_API_KEY for the LangChain ChatOpenAI writer, which defaults to gpt-4.1-mini and changes through LLM_MODEL. Python 3.11+ and uv run the loop. The pytest suite needs no credentials because it scripts model responses, and document access stays on non-symlink .md and .txt files in a single --documents directory.

Do the passing tests prove live routing accuracy?

No. Tests run actual MCP subprocesses and all 12 tools with scripted model responses, require no credentials, and explicitly do not establish Jev's live routing accuracy.

How do I install and run jevex locally?

Requires Python 3.11+ and uv. Run uv sync --group dev, fill TYPESAFE_API_KEY and OPENAI_API_KEY in .env, then run uv run jevex ' ' --trace run.jsonl.

What are jevex's loop limits?

The default loop permits 12 actions, identical calls stop the run, each writer call is capped at 400 output tokens with at most 6,000 characters of observations, and the confidence threshold defaults to 0.35.

Where do Jev's choices and the response go?

Jev's choices and confidence go to stderr and the response goes to stdout. The optional JSONL trace records decisions, arguments, observations and token usage and should be kept private.

CommunityJev Agents & MCP

Jevex

Experiment where Jev controls the tool loop and a LangChain chat model only fills argument values and final prose, using 12 local read-only MCP tools across documents, math, and calendar.

See highlights

Open github.com

Stars: 3
Language: Python
License: MIT
Published by: Community
Author: jvsteiner
Last verified: Oct 2026

jvsteiner/jevex

User repository on GitHub

View on GitHub

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Illustrated architecture explainer — the Jev/LLM split, a live research walkthrough, token accounting, and the limits exposed by the benchmark. A minimal experiment in the architecture described in INSTRUCTIONS.md. Jev directs the loop. A LangChain chat model writes argument values and the final response. Three local MCP servers expose 12 working tools; no external MCP accounts are needed.

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 selects every action and approves exact calls; Python validates arguments and rejected calls never execute.
Three local MCP servers provide 12 read-only tools with no external MCP accounts needed.
JSONL trace records decisions, proposed arguments, observations, and provider token usage for inspection.
Bundled cases read workshop documents then compute answers, such as multiplying prices or measuring date gaps.

Watch out

Prototype limits: 12-action loop, 400-token writer cap, 6,000-character observation bound, read-only tools, and no write operations or production prompt-injection defenses. Document access is limited to non-symlink .md and .txt files in one directory. 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.