Building a local AI browser agent with browser-use and CDP
A first-hand technical report, first published on Tencent Cloud Developer Community, describes adding web-browsing capabilities to a local AI controller called Lingjing using browser-use v0.13.14, Playwright Chromium, and a hybrid CDP-Connect setup. The author documents failures involving Windows subprocess deadlocks, browser startup races, proxy routing and CAPTCHA pages, plus safeguards against high-risk actions, prompt injection, credential exposure, and reliance on unverified third-party data. The author reports that deepseek-chat completed the tested workflow, while the free glm-4-flash option repeatedly failed under the tested version's strict action parsing.
The report provides concrete implementation patterns for controlling Chromium startup, bypassing proxies for local CDP connections, limiting CAPTCHA fallbacks, validating critical data against official sources and timestamps, and guarding browser access. The implementation and results are author-reported and were not independently verified.
技术栈：Python 3.12 + browser-use v0.13.14（MIT）+ Playwright Chromium 本文所有日志片段均来自真机执行记录，未做修饰。 金价数据核实时间：2026-10-30 23:55（GMT+8）
一、为什么需要"双手"
我做的是一个跑在本地的 AI 中控（灵境）。前面几轮迭代下来，它能聊天、能读文件、能查天气、能看图——但一直缺一样东西：它不会上网。
这个缺口比看起来严重。举两个真实例子：
- 我问"今天金价多少"，它只能从训练数据里给我一个印象数字，而那个数字可能是几个月前的；
- 我问"某某公司最近有什么新闻"，它只能说"你可以去搜一下"。
#browser-use#Playwright Chromium#灵境#DeepSeek Chat#GLM-4-Flash#Chromium