jev-ultrafast

Most web agents struggle with latency because they pass raw DOM trees to massive LLMs for every interaction. `jev-ultrafast` takes a different route by pairing TypeSafe with a dynamic, indexed action space to execute browser commands in milliseconds—completing tasks like a Google Flights search from Zürich to London in 7.1 seconds, text generation and page load waits included. ### Architectural Breakdown At the core of `jev-ultrafast` is an observation pipeline that maps page elements directly into a indexed element table: ```text [1] button Change ticket type · Round trip [2] combobox Where from? · San Francisco [3] combobox Where to? · empty [4] textbox Departure · empty ``` Instead of running multiple sequential network calls to pick an action and select a target, the system collapses this into a single TypeSafe request. This request evaluates speculative target questions across specialized target heads simultaneously: * `CLICK` * `TYPE_TEXT` * `SELECT` * `SCROLL_UP` * `SCROLL_DOWN` * `WAIT` * `DONE` * `BLOCKED` The network payload returns the chosen `operation` alongside predictions for `click_target`, `type_text_target`, and `select_target` (if present). The framework then executes the target associated with the selected operation. Target heads only expose compatible elements; for instance, native select options map directly to observed element and option indices. Heavy LLMs are completely decoupled from browser interaction loops. A small LLM is triggered exclusively when the chosen operation is `TYPE_TEXT` to generate the input string. There are no site-specific action scripts or hardcoded field values in the policy. Even the screenshot rendering engine operates out-of-band—it overlays diagnostic labels after decisions are made without driving browser state. ### Setup and Environment Configuration Clone the repository and install dependencies using `uv`: ```bash git clone https://github.com/browser-use/jev-ultrafast.git cd jev-ultrafast uv sync cp .env.example .env ``` Edit your `.env` file to include your required credentials: ```text TYPESAFE_API_KEY=<your_key> TEXT_MODEL_API_KEY=<your_key> ``` Launch the local server: ```bash uv run jev ``` ### Running the Agent and Debug Inspector Once running, open your browser to `http://127.0.0.1:8766`. Chrome connects to the runner via Browser Harness. From the interface, select Start demo → Run automatically. The built-in inspector provides real-time telemetry into the decision-making pipeline, displaying: * Numbered DOM elements from the generated table * Calculated operation probabilities * Target probabilities per head * Executed browser actions If you need to analyze decisions step-by-step, toggle Choose next. This halts execution prior to every browser interaction, allowing you to manually step through each loop iteration.