跳到主要内容
Open In ColabOpen on GitHub

Hyperbrowser 浏览器代理工具

Hyperbrowser 是一个用于运行浏览器代理和扩展无头浏览器的平台。它允许您大规模启动和管理浏览器会话,并为任何网页抓取需求提供易于使用的解决方案,例如抓取单个页面或抓取整个网站。

主要功能

  • 即时可伸缩性 - 在几秒钟内启动数百个浏览器会话,无需基础设施烦恼
  • 简单集成 - 与 Puppeteer 和 Playwright 等流行工具无缝协作
  • 强大的 API - 易于使用的 API,用于抓取/爬取任何网站等
  • 绕过反机器人措施 - 内置隐身模式、广告拦截、自动验证码解决和轮换代理

本笔记本提供了 Hyperbrowser 工具的快速入门概述。

有关 Hyperbrowser 的更多信息,请访问 Hyperbrowser 网站;如果您想查看文档,可以访问 Hyperbrowser 文档

浏览器代理

Hyperbrowser 提供了强大的浏览器代理工具,使 AI 模型能够以编程方式与网页浏览器交互。这些浏览器代理可以导航网站、填写表单、点击按钮、提取数据以及执行复杂的网页自动化任务。

浏览器代理特别适用于

  • 从复杂网站进行网页抓取和数据提取
  • 自动化重复性网页任务
  • 与需要身份验证的 Web 应用程序交互
  • 跨多个网站执行研究
  • 测试 Web 应用程序

Hyperbrowser 提供三种类型的浏览器代理工具

  • 浏览器使用工具:一个通用的浏览器自动化工具
  • OpenAI CUA 工具:与 OpenAI 的计算机使用代理集成
  • Claude 计算机使用工具:与 Anthropic 的 Claude 进行计算机使用集成

概述

集成详情

工具本地可序列化JS 支持
浏览器使用工具langchain-hyperbrowser
OpenAI CUA 工具langchain-hyperbrowser
Claude 计算机使用工具langchain-hyperbrowser

设置

要访问 Hyperbrowser 工具,您需要安装 langchain-hyperbrowser 集成包,并创建一个 Hyperbrowser 账户并获取 API 密钥。

凭证

前往 Hyperbrowser 注册并生成 API 密钥。完成后,请设置 HYPERBROWSER_API_KEY 环境变量

export HYPERBROWSER_API_KEY=<your-api-key>

安装

安装 langchain-hyperbrowser

%pip install -qU langchain-hyperbrowser

实例化

浏览器使用工具

HyperbrowserBrowserUseTool 是一个使用浏览器代理(特别是 Browser-Use 代理)执行网页自动化任务的工具。

from langchain_hyperbrowser import HyperbrowserBrowserUseTool
tool = HyperbrowserBrowserUseTool()

OpenAI CUA 工具

HyperbrowserOpenAICUATool 是一个专门的工具,它通过 Hyperbrowser 利用 OpenAI 的计算机使用代理(CUA)功能。

from langchain_hyperbrowser import HyperbrowserOpenAICUATool
tool = HyperbrowserOpenAICUATool()

Claude 计算机使用工具

HyperbrowserClaudeComputerUseTool 是一个专门的工具,它通过 Hyperbrowser 利用 Claude 的计算机使用功能。

from langchain_hyperbrowser import HyperbrowserClaudeComputerUseTool
tool = HyperbrowserClaudeComputerUseTool()

调用

基本用法

浏览器使用工具

from langchain_hyperbrowser import HyperbrowserBrowserUseTool

tool = HyperbrowserBrowserUseTool()
result = tool.run({"task": "Go to Hacker News and summarize the top 5 posts right now"})
print(result)
{'data': 'The top 5 posts on Hacker News right now are:\n1. Stop Syncing Everything - https://sqlsync.dev/posts/stop-syncing-everything/\n2. Move fast, break things: A review of Abundance by Ezra Klein and Derek Thompson - https://networked.substack.com/p/move-fast-and-break-things\n3. DEDA – Tracking Dots Extraction, Decoding and Anonymisation Toolkit - https://github.com/dfd-tud/deda\n4. Electron band structure in germanium, my ass (2001) - https://pages.cs.wisc.edu/~kovar/hall.html\n5. Show HN: I vibecoded a 35k LoC recipe app - https://www.recipeninja.ai', 'error': None}

OpenAI CUA 工具

from langchain_hyperbrowser import HyperbrowserOpenAICUATool

tool = HyperbrowserOpenAICUATool()
result = tool.run(
{"task": "Go to Hacker News and get me the title of the top 5 posts right now"}
)
print(result)
{'data': 'Here are the titles of the top 5 posts on Hacker News right now:\n\n1. "DEDA – Tracking Dots Extraction, Decoding and Anonymisation Toolkit"\n2. "A man powers home for eight years using a thousand old laptop batteries"\n3. "Electron band structure in Germanium, my ass"\n4. "Bletchley code breaker Betty Webb dies aged 101"\n5. "Show HN: Zig Topological Sort Library for Parallel Processing"', 'error': None}

Claude 计算机使用工具

from langchain_hyperbrowser import HyperbrowserClaudeComputerUseTool

tool = HyperbrowserClaudeComputerUseTool()
result = tool.run({"task": "Go to Hacker News and summarize the top 5 posts right now"})
print(result)
{'data': "Now I'll summarize the top 5 posts on Hacker News as of April 1, 2025:\n\n### Top 5 Hacker News Posts Summary\n\n1. **A man powers home for eight years using a thousand old laptop batteries** (techoreon.com)\n   - 267 points, posted 5 hours ago\n   - An innovative DIY project where someone managed to power their home using recycled laptop batteries for an extended period.\n\n2. **Electron band structure in germanium, my ass** (wisc.edu)\n   - 611 points, posted 8 hours ago\n   - Academic or technical discussion about electron band structure in germanium, possibly with a controversial or humorous take given the title.\n\n3. **Bletchley code breaker Betty Webb dies aged 101** (bbc.com)\n   - 575 points, posted 8 hours ago\n   - Obituary for Betty Webb, who worked as a code breaker at Bletchley Park during WWII, passing away at the age of 101.\n\n4. **Show HN: Zig Topological Sort Library for Parallel Processing** (github.com/williamw520)\n   - 55 points, posted 3 hours ago\n   - A developer sharing a library written in Zig programming language for topological sorting that supports parallel processing.\n\n5. **The Myst Graph: A New Perspective on Myst** (githr.com)\n   - 107 points, posted 5 hours ago\n   - An article presenting a new analysis or visualization of the classic video game Myst, likely using graph theory.\n\nThese are the top 5 posts currently trending on Hacker News as of April 1, 2025.", 'error': None}

带自定义会话选项

所有工具都支持自定义会话选项

result = tool.run(
{
"task": "Go to npmjs.com, and tell me when react package was last updated.",
"session_options": {
"session_options": {"use_proxy": True, "accept_cookies": True}
},
}
)
print(result)
{'data': 'I have found that the react package was last published 11 hours ago. This is the most recently updated package I could find.', 'error': None}

异步用法

所有工具都支持异步使用

async def browse_website():
tool = HyperbrowserBrowserUseTool()
result = await tool.arun(
{
"task": "Go to npmjs.com, click the first visible package, and tell me when it was updated"
}
)
return result


result = await browse_website()
{'data': 'The page displays information about the "Example Domain," stating that it is used for illustrative purposes and can be utilized without permission. There\'s a link to "More information..." but no specific contact details are provided.', 'error': None}

在代理中使用

以下是如何在代理中使用任何 Hyperbrowser 工具的方法

from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_hyperbrowser import browser_use_tool
from langchain_openai import ChatOpenAI
from langgraph.prebuilt import create_react_agent

llm = ChatOpenAI(temperature=0)

# You can use any of the three tools here
browser_use_tool = HyperbrowserBrowserUseTool()
agent = create_react_agent(llm, [browser_use_tool])

user_input = "Go to npmjs.com, and tell me when react package was last updated."
for step in agent.stream(
{"messages": user_input},
stream_mode="values",
):
step["messages"][-1].pretty_print()
================================ Human Message =================================

Go to npmjs.com, and tell me when react package was last updated.
================================== Ai Message ==================================
Tool Calls:
hyperbrowser_browser_use (call_pkAaDjn6kKH9yT3rHDb4hmET)
Call ID: call_pkAaDjn6kKH9yT3rHDb4hmET
Args:
task: Go to npmjs.com and find the last updated date of the React package.
session_options: None
================================= Tool Message =================================
Name: hyperbrowser_browser_use

{"data": "The last updated date of the React package is a day ago.", "error": null}
================================== Ai Message ==================================

The React package was last updated a day ago.

配置选项

Claude Computer Use、OpenAI CUA 和 Browser Use 具有以下可用参数

  • task:要使用代理执行的任务
  • max_steps:代理完成任务所需的最大交互步骤数
  • session_options:浏览器会话配置

更多详情,请参阅相应的 API 参考

API 参考