Agent evals and guardrails as Jev decisions: one cheap, fast request per trace instead of a language-model judge.
@openlayercoWrite an eval once and use it two ways: to test your agent, and to check its actions as it runs. All the checks for a trace go out to Jev as a single request that costs a tiny fraction of a cent.
Works with Jev through the TypeSafe or Vercel APIs, and can fall back to an ordinary chat model if that is all you have, which is slower and costs more.
Installs from PyPI. Built by Openlayer.
Turns any open-weight LLM into a Jev-style decision model, reading typed choices and calibrated probabilities straight from its logits or hidden states.
Jev
Python
vLLMA plugin that lets you ask Jev yes-or-no, pick-one and score questions from Simon Willison's LLM command-line tool.
JevPython
LLMAgent skills and a coding harness where Jev picks the relevant files, so your coding model only reads what matters.
JevPython
Codex
Claude CodeGives Claude Code quick reflexes: Jev scores the possible next steps and hands back the best one.
JevPythonClaude CodeA Claude Code plugin that hides the parts of big tool results you do not need, and brings them back on request.
JevPythonClaude Code
