Jev in Practice: Browser Use and Agent Workflows

Thud uses Jev to decide in-page actions from audio, and is claimed to be 10x faster and 20x cheaper than Gemini Live.

Jev v1.14.0 adds image input, improved multilingual support, and a Box question type that locates objects within an image.

A short video walks through how the image features work in the TypeSafe Jev playground.

A three-layer workflow stores decisions and fixes in files, has Jev score which notes are relevant, and passes the selected context to Opus to continue the work.

The TypeSafe Jev plugin adds image and bounding box support in the Plugin Hub, and existing users can update the plugin.

With the danger zone image and box features enabled, a Sanity Function sets hotspot and crop from a description, which the author calls genuinely useful rather than just a demo.

A link post pointing to the official Jev 1.14 release blog on typesafe.ai, with only the URL and no further usage notes.

Typesafe announced at its Discord town hall that Jev now supports images and a new box primitive, enabled via experimental headers, with the author calling the performance impressive.

Layers Sonnet, Haiku and Opus 5.5 into separate roles, with Jev handling sub-millisecond deterministic branches like file selection and tool dispatch so frontier models only handle genuine forks.

In a tree workflow where Opus writes the plan, Haiku scans and Sonnet patches, the no-thinking micro-branches (which file, which tool, retry or stop) go to Jev for millisecond routing.

Recounts TypeSafe CEO's idea: Jev runs routine decisions, and only hard reasoning wakes Opus with a focused brief before putting it back to sleep.

Jev reads the notes first and Opus only writes what matters; the author claims it cuts weekly chat time by 80% and the vault keeps working offline.