Browser Use
An open-source, event-driven AI agent framework whose core capability is enabling large language models to autonomously control a browser based on natural-language task descriptions. It supports multiple LLM providers and adopts a layered, decoupled architecture that integrates browser control, tool registration, Skills, Sandbox, and MCP.
Core Facts
Browser Use drives browser operations through Playwright at the underlying level, using large language models to understand user intent and incrementally generate execution actions.
Browser Use is an open-source AI web automation agent that allows users to control a browser through natural language instructions.
Browser Use offers two page-understanding modes: DOM mode and visual mode.
Timeline (last 90 days)
Mainstream browser agent frameworks such as Playwright Agent and Browser Use generally adopt a solution combining the accessibility tree with visual screenshots to handle complex UIs.
After installing the BrowserUse plugin, Codex can automatically open a browser, visit websites, click, fill out forms, take screenshots, record screens, and capture information; combined with the computer use plugin, it can control the entire computer
Representative products in the AI browser agent space include Anthropic's Computer Use, Google's Project Mariner, and the independent product Browser Use
Unlike RPA, Browser Use does not rely on pre-recorded fixed scripts; instead, it allows the model to understand page semantics in real time, making it more adaptable to page redesigns.
The Browser Use developer combined Browser Use with Jev to complete automatic ticket booking, with the entire process from opening Google Flights to filtering the cheapest flights taking only about 7 seconds at a cost of approximately $0.0039.
In a flight search case published by Browser Use, the median time using JEF dropped from 9.45 seconds to approximately 7.09 seconds.
The advantage of Browser Use over headless browser scripts such as Selenium and Playwright is that it can leverage existing login sessions without needing to re-handle CAPTCHAs or OAuth flows.
A Twitter comment stated that Browser Use is the best application the author has ever seen.
In the Browser Use combined architecture, Jev is responsible for structured decision-making, another small language model (Mercury 2.5 in the demo) is responsible for generating text input, and the browser code is responsible for executing actions.
Frameworks such as Browser Use and Browserbase encapsulate the LLM call layer, enabling browser operations to be driven by natural language.
6 more timeline events
All Facts (20)
Browser Use drives browser operations through Playwright at the underlying level, using large language models to understand user intent and incrementally generate execution actions.
80%VerifiedBrowser Use is an open-source AI web automation agent that allows users to control a browser through natural language instructions.
80%VerifiedBrowser Use offers two page-understanding modes: DOM mode and visual mode.
80%VerifiedBrowser automation Agents locate webpage elements and simulate human operations through DOM parsing, visual recognition, or Accessibility Tree.
75%VerifiedDeepSeek Harness supports code and file system operations, development task support, collaboration and project planning, and information retrieval capabilities
65%VerifiedCompared to the raw HTML DOM, the Accessibility tree filters out a large amount of purely visual style information and retains semantically interactive elements, reducing token consumption while improving Agent perception accuracy.
65%UnverifiedBrowser Use configuration is divided into three levels: configuration file (.env) → environment variables → API parameters, with priority increasing in that order
100%Partially VerifiedBrowser Use offers three interaction modes: command line (CLI), code invocation (Python API), and online cloud service
100%Partially VerifiedBrowser Use supports connecting to existing browser instances via CDP (Chrome DevTools Protocol) or WebSocket, reusing existing login states and cookies.
100%UnverifiedBrowser Use 采用结构化返回（Structured Output）的方式与大模型交互，而非传统的 Function Calling
95%UnverifiedBrowser Use 内置了十余种常用工具，涵盖搜索、点击、输入、截图等操作，并支持灵活的增删定制
90%UnverifiedBrowser Use 早期版本基于 Playwright 实现浏览器控制，最新版本已切换到 CDP（Chrome DevTools Protocol）
90%UnverifiedBrowser Use官方全面采用异步（async/await）方式编写代码，与Playwright的异步API天然契合
90%UnverifiedBrowser Use的Web UI版本基于Node.js开发，默认运行在7788端口
90%UnverifiedBrowser Use引入LLM作为理解层，不再依赖XPath或CSS选择器等硬编码规则来提取网页信息
85%UnverifiedBrowser Use等大模型驱动工具在每一步操作时需要将当前页面的DOM结构或截图发送给大模型进行推理，一个典型网页DOM序列化后可能超过数万Token
80%Unverified阿里云曾在其技术文章中提到使用 Browser Use 作为核心智能体方案
80%UnverifiedBrowser Use的隔离环境（Sandbox）设计使AI操作的是一个独立的浏览器实例，与用户本机的浏览器会话完全隔离
75%Unverified让同一个 AI 既写代码又测试代码，本质上是自己批改自己的作业，可能系统性忽略认知盲区中的 bug
70%UnverifiedBrowser Use adopts an event-driven layered decoupling architecture, with the core divided into four layers: Agent layer, LLM layer, Tools (Registry) layer, and Browser layer.
60%