I filled in [email protected] and test-password-123, clicked Sign up, and the page showed a welcome message. One image and the favicon failed to load.
What the page shows after submitting
- The heading is still "Join Recipe Club", and the form is still there with both fields filled in.
- A new paragraph reads exactly: "Welcome, [email protected]!". The email is shown in all caps, although I typed it in lowercase.
- Below that is an image with the alt text "logo". It is the one that failed to load (see below).
Resources that failed to load
GET http://127.0.0.1:8765/missing.png(Image) returned 404. This looks like the "logo" image, so it probably shows as broken.GET http://127.0.0.1:8765/favicon.icoreturned 404.- The page document itself (
index.html) loaded with a 200.
I did not check the browser console for JavaScript errors beyond the network requests above.
How it was tried: I installed agent-browser 0.27.0 into a temporary folder (not globally) and ran it with a throwaway home directory, restricted to one local test page I built (a sign-up form with a deliberately missing image and an email-uppercasing handler). Only Skill and agent-browser commands were allowed, with no logins or other sites. The browser download step was not run. I repeated the same steps by hand and got the same page text and the same two 404s. Not tried: logging in, the credential vault, Slack, Electron apps, recording, and any real website.
它做什么
agent-browser 的一个简短发现文件。agent-browser 是为 AI 代理设计的浏览器自动化命令行工具,通过 DevTools 协议驱动 Chrome 或 Chromium,用无障碍树快照和紧凑的元素引用来操作,不依赖 Playwright 或 Puppeteer。
工作方式
- 先安装一次程序:
npm i -g agent-browser && agent-browser install。 - 这个文件只是个入口。执行任何命令之前,代理会用
agent-browser skills get core从命令行里加载真正的使用说明,所以说明始终与已安装的版本一致。 - 另有专门的指南,覆盖 Electron 桌面应用、Slack、探索性测试与 QA、从录制的流量里推导 API 客户端、Vercel Sandbox、受保护的 Vercel 部署,以及 AWS Bedrock AgentCore 云浏览器。
- 可选的可观测性面板运行在 4848 端口。
适合场景
网站自动化、填表、截图、数据提取、网页应用测试。它控制的是真实浏览器,并且能登录网站,请先看风险说明。
高风险:这个包只是一个很短的指引文件。它让代理去全局安装一个 npm 程序(`npm i -g agent-browser && agent-browser install`),并从该程序里加载真正的使用说明,所以实际的指引不在本包里,并且随你安装的版本而变。这个程序能驱动真实的 Chrome:打开网站、点击、填写并提交表单、登录(自带凭据保险库)、在页面上运行 JavaScript、上传文件、录制流量和视频,还能自动化桌面应用和 Slack。浏览器登录了什么,它就能对什么动手。请使用独立的会话或配置,尽量限定可访问的域名,不要在你承受不起误操作的账号上运行,并留意它在做什么。需要 Node.js 和网络。该 Skill 的描述还要求代理优先使用它,而不是内置的浏览器工具。 已试用一次:在隔离环境里(临时家目录、只允许访问 localhost、不登录任何账号)对我自己搭的本地测试页操作,正确完成了填表、提交、读取结果,并列出了加载失败的请求。登录、凭据保险库、Slack、桌面应用、录制和真实网站都没有试,并且模型试用后没有关闭浏览器会话。