Jev Ultrafast is a browser agent that uses a dynamic, indexed action space to complete web tasks with minimal inference overhead.
The tool addresses the problem of slow browser automation by separating decision-making into two lightweight steps: first, the model selects an operation type from those available on the current page, then it selects a target element. This two-decision structure fits into a single network round trip. Text generation happens only when the operation is TYPE_TEXT, avoiding expensive token generation for navigation and interaction tasks. The approach uses a small language model with an OpenAI-compatible interface, supporting providers like OpenRouter, Gemini, GLM, and DeepSeek. Element targeting uses native dropdown indices where applicable, and the action space includes CLICK, TYPE_TEXT, SELECT, SCROLL_UP, SCROLL_DOWN, WAIT, DONE, and BLOCKED operations.
The tool suits developers building web automation systems where latency matters—the demonstration completes a Google Flights search from Zürich to London in 7.1 seconds including actual page loads and text generation. It requires no site-specific scripts or pre-written field strings; the policy works generically across different web interfaces. The project provides a local inspector at http://127.0.0.1:8766 for debugging, with options to step through execution or run automatically. Setup uses uv for dependency management and connects to Chrome through Browser Harness with remote debugging enabled.
The codebase is compact and readable, with the core agent loop available in a single file for inspection. Development appears focused and deliberate, with clear documentation of performance characteristics and a working example that independently verifies outcomes rather than relying on scripted assertions. The project maintains a homepage separate from the repository and provides both a library interface for custom tasks and a complete example implementation.
