Overview & result
Jev Browser combines native Playwright operations with Jev decisions over actual page elements, shipped as a typed SDK, a persistent CLI and an MCP server. Native operations and assertions run without an AI key. Natural-language operations ask Jev to choose among supplied candidates; the README says the decision backend never generates executable JavaScript or selectors.
From the original post
One grounded Jev/Playwright core: typed SDK, persistent CLI, and MCP server with native browser operations and deterministic assertions.
How Jev fits in the loop
- Native Playwright operations and assertions run directly, with no AI key.
- For a natural-language act, observe or extract, the core collects candidate elements from the page's DOM and ARIA tree.
- Jev chooses among the supplied candidates; it never generates JavaScript or selectors.
- Playwright performs the action, and results record how they were verified.
How to reproduce
- Inspect the original showcase and source material at https://github.com/tontoko/jev-browser
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
One core is exposed three ways (typed SDK, persistent CLI, MCP server), and the hosted model can be swapped for any compatible System One endpoint.
Limitations
- The README calls it a functional alternative, not a binary-compatible replacement for the Microsoft or Browserbase packages.
- The AI layer is DOM and ARIA based; Jev does not infer Canvas coordinates or read text from images.
- Hosted services and model weights are not included.