Jev-cu puts a small decision model inside Codex’s UI loop

Jev-cu, a new open-source project from GitHub user Sac-Y, connects TypeSafe AI’s Jev decision model to Codex’s Computer Use runtime. The JavaScript skill assigns Jev the frequent, narrow task of choosing the next interface element, while Codex handles interaction and a local policy gate controls sensitive actions.

Computer-use agents make repeated decisions about which control to select, whether a task is complete, and how risky an action may be. Sending each decision through a large multimodal model can add latency and cost. Jev-cu tests an architecture in which a specialist classifier handles those decisions from text extracted through macOS accessibility APIs.

A classifier in the click loop

Jev returns typed, probabilistic decisions from structured questions and application state. According to TypeSafe AI, it can answer several questions in one parallel pass within 70 to 500 milliseconds. The company reports up to 200 times faster inference and 400 times lower cost than comparable LLMs on classification tasks, with similar performance on the narrow workloads it targets.

Jev-cu uses that interface to select an accessibility candidate, propose an action, report whether the goal is complete, and classify risk. The application can branch on those typed fields without extracting intent from generated prose.

From accessibility text to an action

The repository packages the integration as a Codex skill with a decision loop, policy module, installer, and offline evaluation fixtures. Its current path targets macOS and depends on the accessibility tree, often called the AX tree, which exposes interface controls and labels as structured data.