FAQ

What is Jev, in one sentence?

A model that never writes text: you give it some content and typed questions, and it returns calibrated probabilities over answers you defined.

How is that different from asking an LLM?

An LLM writes a sentence you then have to parse, and it can decide to phrase things differently tomorrow. Jev returns a number in a shape you specified, so your code can branch on it. It is also much faster and much cheaper, because it generates nothing.

It is not a replacement for an LLM. It is the piece you reach for when the job is a decision: classify, route, detect, rate, verify, gate.

What does it cost?

Only input tokens are billed. At the time of writing that is $0.042 per million input tokens for jev-1.13.0, so a short ticket costs a small fraction of a cent. jev always labels a cost as an estimate, and --dry-run prices a run before you send it.

Is this the official TypeSafe CLI?

No. jev is an independent, open-source project and is not affiliated with, endorsed by or sponsored by TypeSafe AI. "TypeSafe" and "Jev" belong to their owner. You need a TypeSafe API key to use it.

Does it send my data anywhere else?

No. jev contacts exactly two hosts: the TypeSafe base URL you configured, when you run a command that evaluates something, and GitHub Releases of this repository, when it checks for an update, which you can turn off. There is no telemetry of any kind. Your state goes to the API and nowhere else, and is never written to a log unless you pass --debug-bodies.

Where is my API key kept?

In TYPESAFE_API_KEY, or in a credentials file with mode 0600 in a 0700 directory. There is no OS keychain, no OAuth and no browser flow, by design. The key is never accepted as a flag value, is held in memory in a type that wipes itself, and only its last four characters are ever shown. See Configuration and profiles.

Are the answers deterministic?

Not bit for bit. Compare against thresholds, never for equality. Pin a versioned model such as jev-1.13.0 for anything repeatable: the alias jev-latest moves to new versions without notice.

What does a probability of 0.5 mean?

That the model cannot tell, not that the answer is "medium". Use --abstain-band 0.4,0.6 to exit 11 for those cases and send them to a person.

What is confidence, then?

How peaked the distribution is, not how likely the answer is to be right. Two genuinely acceptable options give a low confidence, and that is fine.

Can I use it without a key?

Yes, for everything offline: jev validate, jev schema, jev spec, jev completion, jev version and any command with --dry-run. None of them touch the network.

How do I use it in CI?

Put the key in a secret, export it as TYPESAFE_API_KEY, pin a versioned model, and branch on the exit code. CI=true already stops jev prompting and updating itself. See Scripting and CI.

How many rows can jev batch run handle?

As many as you have. Rows are streamed, so memory does not grow with the file, and only ids are kept. Concurrency defaults to 4 and shared keys get rate limited above roughly 8. A crash or a Ctrl-C is not a lost run: --resume picks up from the output file.

Can an AI agent use it?

That is half the point. Every command's help says when to use it, jev spec prints the whole contract as JSON, jev schema prints JSON Schemas, and jev mcp serve exposes the same functionality as MCP tools. See AI agents and MCP.

Which platforms are supported?

Linux (x86-64, arm64), macOS (Intel and Apple silicon) and Windows (x86-64, arm64). One binary, no runtime to install.

Is it stable? Can flags change?

Flags, JSON shapes and exit codes are versioned like an API and follow SemVer. Before 1.0 a breaking change can land in a minor release, and the changelog always says so. Pin a version if you need the contract to hold exactly.

How do I turn off the background update check?

jev config set update.auto false, or JEV_AUTO_UPDATE=false for one environment. It is already off for Homebrew and cargo installs and when CI=true. See Keeping jev up to date.

Something is broken. Where do I report it?

Issues and feature requests: https://github.com/shaharia-lab/jev-cli/issues. Security problems go privately through SECURITY.md, never a public issue. Start with Troubleshooting, which covers most of what goes wrong.