Question 1

Are the 80% token savings real or aspirational?

Accepted Answer

They come from the project's own statistics output, which the README reproduces as a table: response time from 2,400 ms to 890 ms, output tokens from 240 to 48, cost per request from $0.0072 to $0.0018. No prompt set, sample size or error bars are given, and the single console trace shown is one request. Set jev_enabled to false in the config and measure on your own prompts, since the proxy reports per-request Jev latency and agent token counts either way.

Question 2

What stops a misclassification from making the agent do the wrong thing?

Accepted Answer

Nothing enforces the injected verdict, because it is a system-context message rather than a constraint. Category, urgency and action are prepended as plain text and the agent is free to disagree with them. That is also why the README scopes the tool to tasks with clear classification, DevOps prompts and repetitive workflows, and why the proxy is fail-open: a Jev timeout passes the request through instead of blocking your agent.

Question 3

Which agents does it actually support?

Accepted Answer

Claude Code, Codex, Kilo Code and Cline, detected automatically, each with its own adapter covered by tests. Wiring is a base URL: ANTHROPIC_BASE_URL for Claude Code and Kilo, OPENAI_BASE_URL for Codex, pointed at the local proxy. Streaming SSE with token extraction, multi-turn conversation context and media bypass for images and PDFs are handled in the proxy, and config location and port are overridable through JEVXAGENT_CONFIG and JEVXAGENT_PORT.

Question 4

What happens to my API key and prompt traffic?

Accepted Answer

The key is entered during interactive setup and written to ~/.jevxagent/config.json; the README states there is no credential logging and that the proxy binds to localhost only. The trade-off is that every prompt is classified by Jev before your agent sees it, so if that classification is not something you want leaving the machine this design offers no offline mode. Version 0.2, MIT licensed, with a live console display and JSONL statistics for per-request inspection.

Question 5

How do I start the proxy and point my agent at it?

Accepted Answer

Clone the repo, run pip install -e ., then run python -m jevxagent and answer the interactive setup prompts. Set ANTHROPIC_BASE_URL for Claude Code or Kilo, or OPENAI_BASE_URL for Codex, to the local proxy address, which defaults to http://127.0.0.1:9099.

Question 6

What does the Line J injection actually add?

Accepted Answer

A system message carrying Jev's category, urgency and action, placed ahead of the user prompt, so the agent sees what was already decided before it generates. Images and PDFs skip Jev automatically and pass straight through.

Question 7

Where do the configuration and statistics live?

Accepted Answer

Config is written to ~/.jevxagent/config.json, overridable with JEVXAGENT_CONFIG, and every request is appended to a JSONL statistics file that can be kept per-project. Ctrl+S prints a cumulative snapshot, and JEVXAGENT_PORT changes the proxy port.

jevXagent

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community