面向AI代理的开源浏览器

theredsix · 2026-03-11T14:39:30.000Z

Hi HN, I forked chromium and built agent-browser-protocol (ABP) after noticing that most browser-agent failures aren’t really about the model misunderstanding the page. Instead, the problem is that the model is reasoning from a stale state. ABP is designed to keep the acting agent synchronized with the browser at every step. After each action (click, type, etc), it freezes JavaScript execution and rendering, then captures the resulting state. It also compiles the notable events that occurred during that action loop, such as navigation, file pickers, permission prompts, alerts, and downloads, and sends that along with a screenshot of the frozen page state back to the agent. The result is that browser interaction starts to feel more like a multimodal chat loop. The agent takes an action, gets back a fresh visual state and a structured summary of what happened, then decides what to do next from there. That fits much better with how LLMs already work. A few common browser-use failures ABP helps eliminate: A modal appears after the last Playwright screenshot and blocks the input the agent was about to use Dynamic filters cause the page to reflow between steps An autocomplete dropdown opens and covers the element the agent intended to click alert / confirm interrupts the flow * Downloads are triggered, but the agent has no reliable way to know when they’ve completed As proof, ABP with opus 4.6 as the driver scores 90.5% on the Online Mind2Web benchmark. I think modern LLMs already understand websites, they just need a better tool to interact with them. Happy to answer questions about the architecture, forking chrome or anything else in the comments below. Try it out: claude mcp add browser -- npx -y agent-browser-protocol --mcp (Codex/OpenCode instructions in the docs) Demo video:

theredsix，持续分享值得阅读的内容。关注

你好，HN，我fork了chromium，并构建了代理-浏览器协议（ABP），因为我注意到大多数浏览器代理失败其实并不是模型误解页面。相反，问题在于模型是在陈旧状态下推理。

ABP旨在确保执行代理在每一步都与浏览器同步。每次操作（点击、输入等）后，它会冻结JavaScript的执行和渲染，然后捕获生成的状态。

它还会汇总该动作循环中发生的显著事件，如导航、文件选择器、权限提示、警报和下载，并连同冻结页面状态的截图一起发送给代理。

结果是，浏览器交互开始更像是多模态聊天循环。代理采取行动后，获得新的视觉状态和结构化的事件摘要，然后决定下一步行动。这更符合大型语言模型的工作原理。

ABP帮助消除的几个常见浏览器使用失败：最后一张Playwright截图后出现模态，阻断代理即将使用的输入;动态过滤器使页面在步骤间重流;自动补全下拉菜单打开并覆盖代理计划的元素;alert / confirm会中断流程 * 下载已触发，但代理无法可靠知道他们何时完成了作为证明，ABP 作为 Opus 4.6 的驾驶员在 Online Mind2Web 基准测试中得分为 90.5%。

我认为现代大型语言模型已经理解网站，只是需要更好的工具来与网站互动。欢迎在下方评论区回答关于架构、分支Chrome或其他任何问题。

试试看：claude mcp add browser -- npx -y agent-browser-protocol --mcp（Codex/OpenCode 说明见文档）演示视频：

添加评论

点踩

分享查看原文

参与讨论