Question 1

Is this just curl with nicer ergonomics, or does it stop me wasting calls?

Accepted Answer

Mostly validation. The API accepts and charges for questions that cannot inform you, all returning 200: a score with one level comes back as score 0 with confidence 1, and a choice with one option returns it at probability 1.0. The dangerous case is criteria that restate the label, where the result still looks fine, so the offline rules check before you spend.

Question 2

Do the terse YAML question sets actually cost less, or just write less?

Accepted Answer

Just write less. The terse form is 198 characters against 308 raw, a 36% reduction in what you type, and it expands back into the same JSON before sending, so the wire payload and the API cost are identical. Nothing in the README claims a cheaper call.

Question 3

Will it make my calls faster than hitting the API myself?

Accepted Answer

No, and the tool does not claim it does: the API is around 273 ms per call and nothing makes ask faster than the network. What it adds is offline linting in roughly 0.5 ms with no network call at all, fast enough for a pre-commit hook, plus batching, since questions in one request are answered in parallel and cost more without adding latency.

Question 4

Where does my key live, and does this run on Windows?

Accepted Answer

Keys sit in the OS credential store, never argv or shell history, and resolution order is --api-key, then the provider environment variable, then the keyring. auth login prompts on a hidden TTY and never takes a key as an argument. Prebuilt binaries cover Linux and macOS on x86_64 and aarch64; Windows needs WSL.

Question 5

How does it behave in CI, and what do the exit codes mean?

Accepted Answer

The codes are a documented contract: 0 for success, 1 for input, credential, or API failure, and 2 for lint errors or invalid arguments, with 2 also meaning warnings only under jev lint. Linting runs automatically on ask, errors block the call and warnings print to stderr, and --strict or --deny-warnings makes warnings fail too. --json gives machine-readable findings.

Question 6

Can a coding agent drive it, and how does the skill install?

Accepted Answer

Yes. The repo ships an agent skill in .skills/jevkit/ that teaches a coding agent the CLI and the failure modes that are easy to hit, and npx skills add ariel-frischer/jevkit registers it for Claude Code, Codex, Gemini CLI, OpenCode, and Zed at once, since they all read ~/.agents/skills. Confirm it with npx skills ls -g.

Question 7

Can I see what a call cost, or point jevkit at a proxy?

Accepted Answer

Both. jev ask --log appends one JSON line per call with request, response, tokens, and cost to ~/.local/state/jev/usage.jsonl, and --raw prints the full response with usage and cost. The endpoint key in ~/.config/jev/config.toml can redirect to a proxy or a self-hosted install, with settings layered flags over JEV_* environment variables over the config file over defaults.

Jevkit — ariel-frischer

About this project

Highlights

Quickstart

Questions about this entry

From the Jev guides

From the Jev AI model community