What is Jev Ultrafast?
Jev Ultrafast is an ultra-high-speed autonomous web agent framework created through a collaboration between Browser Use (the creators of the popular open-source browser automation agent) and TypeSafe. Engineered to resolve the notorious latency and fragility of traditional vision-based browser agents, Jev Ultrafast introduces a dynamic, indexed action space combined with a dual-head decision policy. In live real-world benchmarks, the agent completes an end-to-end flight search workflow on Google Flights (Zürich to London) in just 7.07 seconds at 1× real-time speed, including live text generation, dynamic dropdown selection, and full page loading waits.
Key Features & Technical Innovations
- Dynamic Indexed Action Space: Converts real-time DOM observations into structured, numbered element tables (e.g.
[1] button,[2] combobox,[3] textbox), ensuring the model only acts upon valid, actionable targets. - Single-Round-Trip Dual-Head Decisions: Powered by TypeSafe Jev. A single model request predicts both the operation (
CLICK,TYPE_TEXT,SELECT,SCROLL_UP,SCROLL_DOWN,WAIT,DONE,BLOCKED) and the speculative target index simultaneously, cutting network round trips in half. - Decoupled Asynchronous Text Generation: Text input is generated only when the chosen operation is
TYPE_TEXT, using an ultra-fast compact LLM (such as Mercury 2.5, DeepSeek, GLM, or Gemini), eliminating heavy token generation overhead during click navigation. - 90.7% Reduction in Protocol Overhead: Optimized DOM snapshotting and event-driven combobox waiting reduces browser protocol calls from 1,092 down to just 101 per task, slashing median Jev decision latency to 178 ms.
- Visual Inspector & Debugging UI: Built-in local web dashboard at
http://127.0.0.1:8766displaying element tables, operation probabilities, target distributions, and step-by-step action replays. - Robust Chrome CDP Connection: Connects natively via Browser Harness with full remote debugging support, cookie persistence, and headless or headed browser execution.
Benchmark: Live Google Flights Workflow Performance
| Performance Metric | Conventional Browser Agent | Jev Ultrafast (Optimized) | Improvement |
|---|---|---|---|
| Task Execution Duration | 11.21 s | 7.07 s | 25–37% faster |
| Median Jev Decision Latency | 450+ ms | 178 ms | 60% latency reduction |
| Browser Protocol Calls (CDP) | 1,092 calls | 101 calls | 90.7% overhead reduction |
| City String Generation | Included in prompt loop | 346–581 ms (Isolated) | Zero blocking on click steps |
| Decision Round Trips | 2 round trips per step | 1 round trip (Dual-head) | 50% API round-trip saving |
Quickstart & Local Setup
- Clone & Install Dependencies:
git clone https://github.com/browser-use/jev-ultrafast.git cd jev-ultrafast uv sync - Configure Credentials: Copy
.env.exampleto.envand add yourTYPESAFE_API_KEYandTEXT_MODEL_API_KEY(OpenRouter, Gemini, or DeepSeek). - Launch the Visual Inspector:
uv run jevOpen
http://127.0.0.1:8766to watch autonomous navigation in real time.
Python Library Integration
from jev_ultrafast import Agent
with Agent(
"https://www.google.com/travel/flights?hl=en",
"Find one-way flights from Zurich to London on September 20, 2026, for one adult in economy."
) as agent:
agent.run()
Technical Specifications
| Core Framework | Python / uv / Chrome DevTools Protocol (CDP) |
|---|---|
| Decision Model | TypeSafe Jev (Dual-head operation + target) |
| Text Helper Models | Mercury 2.5, DeepSeek V3, Gemini 2.0, GLM-4 |
| Browser Automation | Browser Harness / Native Google Chrome |
| License | Open Source (Apache-2.0 / MIT compatible) |
| Official Repository | https://github.com/browser-use/jev-ultrafast |
| Parent Organization | Browser Use (50k+ GitHub Stars) |
Frequently Asked Questions
What makes Jev Ultrafast faster than standard vision-based agents?
Rather than feeding massive screenshots into multimodal LLMs on every single frame, Jev converts actionable controls into a structured indexed element table and resolves both action type and target element in a single 178ms network request.
Can Jev Ultrafast handle dynamic web interfaces like comboboxes and calendars?
Yes. The engine includes event-based combobox synchronization and dynamic element tables that refresh after state mutations, allowing it to navigate complex booking engines, popups, and datepickers reliably.
Which LLM providers are supported for text inputs?
Any OpenAI-compatible text endpoint can be configured as the text helper, including OpenRouter (Mercury 2.5), Google Gemini, DeepSeek, and local Ollama models.
Is Jev Ultrafast free to use?
The core library and browser automation harness are completely open source on GitHub. Users only need API keys for the TypeSafe policy model and chosen text helper LLM.