Overview & result
Browser automation as a library, CLI and MCP server. The calling LLM states the outcome it wants one step at a time, code describes the page, and a single System One request asks Jev which element, which action, which value, and whether the step is done, blocked, showing an error or about to do something irreversible. Playwright performs the action. The author describes each request as taking about 300 ms.
From the original post
Browser automation where an LLM plans and Jev (Typesafe System One) decides. Library, CLI and MCP server.
How Jev fits in the loop
- The calling LLM states the outcome it wants for one step and supplies any text to type.
- Code describes the current page.
- One System One request asks Jev which element, which action and which value, and whether the step is done, blocked, showing an error or irreversible.
- Playwright performs the action; the LLM does not read page snapshots unless it takes over.
How to reproduce
- Inspect the original showcase and source material at https://github.com/Ying-Kai-Liao/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
The author reports 40 of 42 tasks correct on live sites with no false "done" claims, and far less page content sent to the LLM than a Playwright-MCP-style loop.
Limitations
- Results are the author's own runs; the token comparison is labelled an estimate.
- Remaining misses are counting tasks and verifying a sort.
- Unofficial project that uses your own TypeSafe API key.