Question 1

Is this just a bash wrapper around the Jev API?

Accepted Answer

It is a wrapper, and the interesting part is the shape of the calls. Five tools, jev_classify, jev_check, jev_score, jev_rank, and jev_ask, expose the same names and the same JSON in every harness. Each validates input locally (shapes, duplicate ids, request size) before spending a call, batches all items into one request, and returns auto or review decisions computed from thresholds you can override per call.

Question 2

Will this rewrite my agent config files?

Accepted Answer

Yes, that is the install path, and you can preview it. `npx -y @french-castle/jev-code@latest setup` detects Claude Code, Codex, Pi, and OpenCode, copies the skill and registers the tool in each, and asks for a key with hidden input when none is in your shell. --dry-run prints the plan, --project scopes to the current repo instead of your user profile, and the OpenCode config is backed up before editing.

Question 3

Can I use it in a harness with no MCP client?

Accepted Answer

Yes. A CLI exposes the same tools and JSON, which is what lets the skill work where no tool is registered, and the skill alone installs into 70-plus agents through the skills.sh installer if you would rather wire up MCP yourself. Pi has no MCP client at all, so there the five tools are registered natively as a pi extension and you run /reload afterwards.

Question 4

Which host does my key belong to, and how do I confirm it works?

Accepted Answer

The key prefix picks the host: ts_ for TypeSafe, sk-or- for OpenRouter, vck_ for Vercel AI Gateway. `doctor` reports which host and key are in use and what is installed, and `doctor --live` makes a real call. Exporting the key beforehand is picked up without a prompt, and setup prints the line to add to your shell profile so Pi and the CLI read the same value.

Question 5

What are the five tools and what does each return?

Accepted Answer

jev_classify gives a label, probabilities, margin and an auto or review decision; jev_check gives a probability and a verdict of yes, no or uncertain; jev_score gives a score, nearest level, confidence and decision; jev_rank gives relevance per candidate sorted plus any_relevant; and jev_ask returns raw System One answers.

Question 6

How does setup handle my agent's config files?

Accepted Answer

npx @french-castle/jev-code@latest setup detects Claude Code, Codex, Pi and OpenCode, copies the skill and registers the tool in each. --dry-run prints the plan first, --project scopes it to the current repo, and the OpenCode config is backed up before editing. Pi registers natively and the others over MCP stdio.

Question 7

Can the skill work in an agent with no MCP tool registered?

Accepted Answer

Yes. A CLI exposes the same tools and the same JSON, and the skill is a standard Agent Skills directory, so the skills.sh installer adds it to 70-plus agents. You pair it with the MCP server or let the agent fall back to the CLI.

Jev Code · FrancoisChastel

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community