Question 1

Why pay for a second model call on every message and step?

Accepted Answer

Jev returns typed answers with confidence in about 70 to 500 ms for roughly a ten-thousandth of a dollar per call. Generation, file edits, and tool execution stay with the main model; Jev only picks the route.

Question 2

Are the routing accuracy numbers real or demo-sized?

Accepted Answer

Skill routing hit 85.9% with zero wrong picks on 64 requests over 22 skills, and 75% on a 78-skill Hermes roster. The verification gate scored 80% on 15 cases but ships off pending a larger run.

Question 3

What breaks if Jev is slow, wrong, or unreachable mid-session?

Accepted Answer

Install with bunx install or the native plugin entry plus OPENROUTER_API_KEY, then tune gate thresholds, max tools, state budget, and timeouts. Low-confidence or malformed answers change nothing and never block.

Question 4

What does browser_task actually send and require to run?

Accepted Answer

It also sends page text and controls each step while running, needs uv plus a browser over CDP, and only TYPE_TEXT calls a text model. The documented live run took 2 steps and 3 decisions on the correct page.

Question 5

How do I install it, and what does it touch on my machine?

Accepted Answer

bunx jev-for-all@latest init, or npx if you prefer, and use @latest because both cache the first version they ever fetched. It needs bun, and it does four things: writes ~/.config/jev-for-all/config.json with your OpenRouter key at mode 0600 plus a models.json of tiers, registers the OpenCode plugin, installs the Claude Code plugin from this repo's marketplace, and warns if the OpenCode Go login is missing. Then jev-for-all doctor checks every piece.

Question 6

Where does my text go, and what is the local alternative?

Accepted Answer

Decisions are made by Jev through OpenRouter's alpha Decisions API, which the README notes can change or disappear. Your prompt goes there, and in the OpenCode plugin the recent conversation goes too, including tool output. The alternative is a local Laya URL passed with --laya, which is free but, in the author's words, much less accurate than Jev.

Question 7

Should I trust this on anything that matters?

Accepted Answer

The README says no, plainly: it is fully experimental and should not go into important or production-grade projects. It does not edit your code, your git history or your agent settings, and every failure path fails open rather than blocking you. But the evidence is thin, 12 small single-file tasks written by the author, and in his own tests it sent about 2 of every 58 ordinary prompts to the cheaper tier when a stronger one was the better fit.

Jev For All

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev community