Overview & Result
Hrishikesh Chappadi published jev-browser-wingman, an open-source tool that lets coding agents use a real logged-in Chrome session. The large model keeps planning while Jev selects what to click or type at each bounded step.
How Jev fits in the loop
- Claude Code or OpenCode keeps the task plan and text generation.
- The visible Chrome state is serialized into a bounded decision and Jev chooses the next click or type operation.
- The local driver executes one step and returns a typed status such as done, blocked, ambiguous, or fallback.
How to reproduce
- Clone the verified repository and read its agent installation instructions.
- Attach the tool to an already-open Chrome session as documented by the project.
- Pass a bounded browser goal and inspect the returned action status before allowing the next step.
Why this build matters
Repeated browser micro-actions become a typed decision loop instead of a full generative turn.
Limitations
- The author states that the tool is not affiliated with TypeSafe.
- It is intentionally bounded: it does not navigate, open tabs, sign in, or generate the agent's text.