The official post says developers can pass a model, a state (a string or JSON), and questions such as a choice, score, or probability estimate; responses are structured around the types and choices defined by the caller. The illustrated setup uses uv add ai and an AI Gateway key.
The post presents Jev as suited to narrow classification decisions rather than open-ended text generation. Its author reports that Jev performed better than a hand-built classifier on a Python-versus-English REPL use case, but still misclassified a partially typed expression. A separate experiment built Python code by choosing AST nodes; the resulting code was syntactically valid but mostly incorrect. Vercel describes the API as experimental and says users should test whether Jev is accurate enough for their own use cases.