Why I Moved the Click Decisions Out of the LLM: jev-ultrafast and a Local AI Gat
说明与用途
该方案将快速的Jev模型用于点击决策,通用LLM用于规划,ServBay本地AI网关负责请求路由,实现快速可靠的浏览器自动化。
问题
General-purpose LLMs are slow and fragile for frequent click interactions in browser automation workflows.
结果
The combined setup completes a Google Flights search workflow in ~7 s, with micro‑clicks handled in tens of milliseconds and macro‑planning by a general LLM, while local API routing eliminates external latency and improves resilience.
工作流
- Clone jev-ultrafast repo
- Sync dependencies with ServBay’s Python environment using uv
- Install Playwright Chromium
- Run browser-harness --doctor to verify Chrome debugging
- Add OpenAI/Anthropic/DeepSeek channel in ServBay AI Gateway
- Create project‑level virtual key
- Configure ServBay to map OpenAI SDK calls to chosen provider
- Use Jev model for micro‑click decisions and LLM for macro‑planning
- Reuse local Chrome login state to bypass 2FA
- Provision upstream channels automatically via Python script
证据
jev-ultrafast passed 21,000 stars on GitHub shortly after the Browser-Use team open sourced it, and the reason is not hard to find. It never asks a general purpose model to write a paragraph of analysis before a click.
In tests the project completes the whole Google Flights search and comparison workflow in about 7 seconds on its own, including dynamic dropdown menus, multi-step interaction wizards and overlay pop-ups that appear without warning.
The setup I ended up with pairs the local AI gateway in ServBay, which runs on macOS and Windows, with the human‑like interaction model from Browser‑Use.
- 记录日期: 2026-10-12T16:41:13.264Z