Overview & result
The project gives Jev a bounded set of page actions and lets it pick one step at a time while code controls budgets, recovery, and stop conditions.
From the original post
Fast and very cheap browser use using TypeSafe's Jev model.
Give jev-browser a task and a URL.
It drives a real headless browser through an MCP server, CLI, or library. TypeSafe's Jev model picks one action per step from the page's clickable, typeable, and selectable elements, and scores how likely it is that the goal is met or the run is stuck. Code owns the loop: budgets, recovery, stop gates. You get the final page, a step trace with confidences, console errors, and a screenshot.
How Jev fits in the loop
- Turn clickable, typeable, and selectable elements into candidate actions.
- Ask Jev to choose the next action.
- Score whether the goal is complete or the run is stuck.
- Keep control flow and safety budgets in ordinary code.
How to reproduce
- Clone the repository and follow the author's current setup instructions.
- Provide a TypeSafe API key locally; never expose it in client-side code or commit it to GitHub.
- Run a bounded browser task, record the page states and choices, then compare cost, latency, completion, and failure behavior.
Why this build matters
It demonstrates the exact thesis behind JevForAgents: Jev can be the decision layer inside an agent loop while deterministic code owns execution.
Reported performance
Reported by authorCost: $0.0016 for one reported Wikipedia navigation example
Latency: ~4s end-to-end for that reported example
Limitations
- The demo and performance figures come from the project author.
- Browser pages change, so element selection and task completion can regress.
- Sensitive answers and irreversible actions should remain behind explicit user confirmation.