Question 1

Is this a thin wrapper around one vendor's model?

Accepted Answer

No. The same binary talks to anything behind the System One API: Jev by default, d1 through Liquid AI's hosted endpoint, and self-hosted CLM or Laya by pointing TYPESAFE_BASE_URL at your own server. Coverage differs per model, though: Span-01 answers noul questions only.

Question 2

Will one bad row blow up a 500-row batch?

Accepted Answer

No. One call accepts up to 500 records as items, and a record that fails does not fail the rest. Questions within a call run in parallel, which is also why a single call can cheaply cover urgency, ownership and sentiment for the same input.

Question 3

Can I point it at a decision model running on my own machine?

Accepted Answer

Yes. Set TYPESAFE_BASE_URL at a local Laya or CLM server and keep TYPESAFE_API_KEY set, because that variable is what selects the route; any non-empty value works if your server ignores keys. Profiles let you switch models without rerunning setup.

Question 4

What does the installer actually do to my agent setup?

Accepted Answer

It writes MCP registration. evaluate setup mcp finds Claude Code, Claude Desktop, Codex and Hermes and registers the evaluate tool with each; pi needs its own setup subcommand. The install path is documented for macOS and Linux as one static binary with no Node or Python runtime, and evaluate update upgrades it in place.

Question 5

How do I install it and connect my agents?

Accepted Answer

macOS and Linux use a one-line install script, then TYPESAFE_API_KEY=your-key evaluate setup mcp. It finds Claude Code, Claude Desktop, Codex and Hermes and registers evaluate with each, and re-running setup updates them. An OpenRouter account can provide the key instead.

Question 6

What does one evaluate call actually return?

Accepted Answer

Typed answers an agent can branch on: noul gives a probability, choice gives the chosen option with a confidence and a probability per option, and score rates on a scale you define. The worked example returns is_urgent 0.95 and department billing with 86% probability.

Question 7

How fast is it, and can it handle many questions at once?

Accepted Answer

Jev typically answers in under half a second. One call can ask several questions, and they run in parallel. You can also pass up to 500 records as items and ask the same questions of each, and if one record fails the rest still complete. It is a single static binary with no Node or Python runtime.

System One Connector

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community