Open Source 开源项目 4d ago Updated 4d ago 更新于 4天前 59

browser-use/browser-use browser-use/browser-use

Browser Use is an open-source AI agent library that enables LLMs to control web browsers like humans — clicking, typing, scrolling, and filling forms It ranks #1 on the Odysseys leaderboard with 87.4% average across 200 long-horizon web tasks, outperforming agents from OpenAI, Anthropic, Google, and Microsoft Offers two deployment modes: a free open-source Python library (runs locally) and a fully-hosted cloud agent with built-in stealth, proxy rotation, and captcha solving The optimized `ChatBr Browser Use 是一个开源 AI 浏览器自动化框架,允许 LLM 像人类一样操作网页(点击、填表、提取数据等),支持 Python 库和 CLI 两种集成方式 在 Odysseys 排行榜(200 个长周期网页任务)以 87.4% 平均分位居第一,领先 OpenAI、Anthropic、Google、Microsoft 的计算机使用代理 提供自托管开源版本和全托管云服务两种模式,云服务内置代理轮换、验证码解决和 1000+ 集成(Gmail/Slack/Notion 等) 支持多 LLM 后端(OpenAI/Anthropic/Google 及自有优化模型),通过单一 API Key

70
Hot 热度
65
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Browser Use is an open-source AI agent library that enables LLMs to control web browsers like humans — clicking, typing, scrolling, and filling forms
  • It ranks #1 on the Odysseys leaderboard with 87.4% average across 200 long-horizon web tasks, outperforming agents from OpenAI, Anthropic, Google, and Microsoft
  • Offers two deployment modes: a free open-source Python library (runs locally) and a fully-hosted cloud agent with built-in stealth, proxy rotation, and captcha solving
  • The optimized ChatBrowserUse() model completes browser tasks 3-5x faster than other models with state-of-the-art accuracy
  • Supports any LLM provider through a unified API key, including OpenAI, Anthropic, Google, and Browser Use's own bu-* models

Why It Matters

Browser Use addresses a critical gap in AI agent capabilities — reliable, autonomous web interaction at scale. For AI practitioners building agents that need to interact with real-world web applications (scraping, automation, research), it provides a battle-tested framework that outperforms major tech companies' offerings on long-horizon tasks.

Technical Details

  • Architecture: Python-based agent framework (requires Python >= 3.11) with deep code-level integration, allowing users to pick their LLM, customize system prompts, and add custom tools
  • Two modes: Open-source agent (free, runs locally with full control) and fully-hosted cloud agent (recommended for complex tasks, includes 1000+ integrations like Gmail/Slack/Notion, persistent filesystem, and memory)
  • Benchmarking: Evaluated across 100 real-world browser tasks (open-source benchmark at browser-use/benchmark) and the Odysseys leaderboard (200 long-horizon web tasks)
  • LLM flexibility: ChatBrowserUse accepts provider-prefixed model IDs (e.g., anthropic/claude-sonnet-4-6, openai/gpt-5.5, google/gemini-3-pro) through a single BROWSER_USE_API_KEY
  • CLI vs Library: CLI mode for one-off tasks through existing agents (Claude Code, Cursor, etc.); Python library for repeatable automation, scheduling, parallel execution, and embedding into custom products

Industry Insight

  • The dominance on the Odysseys leaderboard suggests that specialized optimization for browser automation (rather than general-purpose agents) yields significantly better results — a signal for the industry to invest in task-specific agent tuning
  • The dual deployment model (open-source + cloud) reflects a growing pattern where AI tooling companies compete on both accessibility and managed infrastructure, making it easier for teams to prototype locally and scale in production
  • The unified API key abstraction across LLM providers reduces integration friction and could accelerate adoption of browser agents in enterprise workflows where multi-provider strategies are common

TL;DR

  • Browser Use 是一个开源 AI 浏览器自动化框架,允许 LLM 像人类一样操作网页(点击、填表、提取数据等),支持 Python 库和 CLI 两种集成方式
  • 在 Odysseys 排行榜(200 个长周期网页任务)以 87.4% 平均分位居第一,领先 OpenAI、Anthropic、Google、Microsoft 的计算机使用代理
  • 提供自托管开源版本和全托管云服务两种模式,云服务内置代理轮换、验证码解决和 1000+ 集成(Gmail/Slack/Notion 等)
  • 支持多 LLM 后端(OpenAI/Anthropic/Google 及自有优化模型),通过单一 API Key 统一调用,推荐模型完成速度比通用模型快 3-5 倍
  • 基准测试覆盖 100 个真实网页任务,代码和评测数据完全开源

为什么值得看

Browser Use 在浏览器自动化 Agent 赛道取得性能领先,证明了"LLM+浏览器控制"路线的工程可行性,为开发者提供了从原型到生产的一站式解决方案。其开源核心与云服务并行的模式降低了 AI 网页自动化门槛,对需要批量网页操作、RPA 替代或 Agent 集成的团队具有直接参考价值。

技术解析

  • 架构设计:基于 Python 的 Agent 类封装浏览器操作,通过 ChatBrowserUse 抽象层统一对接不同 LLM 提供商,支持自定义工具扩展(@tools.action 装饰器)和系统提示词覆盖
  • 模型优化:自有 bu-* 系列模型针对浏览器任务微调,在速度和准确率上显著优于通用模型;同时兼容 OpenAI/Anthropic/Google 模型,通过 provider/model-id 格式统一路由
  • 双模式交付:开源版提供代码级控制(适合嵌入产品、定时任务、并行 scraping);云服务版提供托管浏览器、持久化存储和免维护基础设施(适合复杂长周期任务)
  • 基准验证:在 100 个真实网页任务基准和 200 任务 Odysseys 排行榜上公开评测,开源评测代码位于 browser-use/benchmark 仓库
  • 集成生态:CLI 版本可无缝接入 Claude Code、Cursor、Codex 等现有 Agent 工作流,Python 库支持结构化输出、细粒度浏览器控制和自定义系统提示

行业启示

  • 浏览器自动化进入 Agent 原生时代:传统 RPA 工具正被 LLM 驱动的语义理解能力替代,Browser Use 的性能领先表明该赛道已出现明确的技术标杆,后续竞争将聚焦于长周期任务稳定性和复杂交互场景覆盖
  • 多 LLM 兼容成为基础设施标配:通过统一 API Key 聚合多家模型供应商,既降低集成成本又避免厂商锁定,这种"模型中立"架构将成为 AI Agent 平台的通用设计模式
  • 开源核心+云服务变现的平衡策略:开源版本建立生态和信任,云服务提供开箱即用的企业级能力(代理轮换、验证码解决、持久化),这种分层交付模式可复制到其他 Agent 工具领域

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 Agent Agent Programming 编程