Overview & result
Fahad Saleem describes jev-browse as a one-call handoff for website sub-tasks. A small typed Jev model chooses clicks and typing in a background tab, while the main Claude Code or Codex agent returns only when the task is stuck, risky, or complete.
From the original post
When your coding agent drives Chrome itself, every click burns a full model turn. A five-field form becomes fifteen expensive turns of re-reading the whole context.
jev-browse hands the website sub-task off in one call:
• `fast_run(url, goal)` opens a background tab
• a small typed model (Jev) picks each click/type in a few hundred ms
• your agent only gets pulled back when something is stuck, risky, or done
• form and nav work costs far fewer agent turns than driving the browser yourself
Works with Claude Code and Codex. Needs Chrome, browser-harness, and a TypeSafe key.
https://t.co/uk9P9R2gs4
How Jev fits in the loop
- The coding agent supplies a URL and bounded goal through fast_run(url, goal).
- Jev selects the next browser click or type action in the background tab.
- The helper returns a typed outcome to the agent only at a stop condition or handoff point.
How to reproduce
- Clone the verified repository and follow its browser-harness installation and doctor checks.
- Call the documented fast_run(url, goal) helper against a non-sensitive test page.
- Inspect the typed completion, stuck, risky, or done result before handing control back to the coding agent.
Why this build matters
Keep planning in the coding agent and move repetitive, low-level browser choices into Jev.
Reported performance
Latency: A few hundred ms per Jev decision (author-reported)
Limitations
- The project requires Chrome, browser-harness, and a TypeSafe key.
- The author describes the latency and turn reduction; the post does not provide a controlled baseline table.