Observe
Read pages, tabs, element outlines and current window state.
A real browser workspace for AI Agents.
快速了解 AgentBrowser 能做什么、哪些步骤需要人工确认，以及如何接入现有 AI 工作流。
Architecture
Browser, protocol and authorization are separate layers.
从页面观察、权限判断到真实操作与Human handoff，每一步都有明确边界。
Authorize
Authorize advanced actions, downloads and window control first.
Act
Log clicks, input, scrolling, tab switches and downloads in the protocol.
Handoff
Let users visibly complete sign-in, CAPTCHAs and payment.
Agent integration example
Any Agent framework that can send HTTP requests can connect.
POST /agent/v2/act
{ "action": "click", "selector": "#submit-btn" }
→ 200 OK
{ "ok": true, "result": "clicked" }
LangChain / AutoGen / Custom Agent frameworks connect over HTTP; use Human Handoff for sign-in, CAPTCHAs and payments.
Window control
Maximize, minimize, bring forward, capture screenshots and switch tabs.
Download distribution
Publish installers and updates centrally to avoid version confusion.
Audit records
Record key actions in protocol logs for diagnosis and testing.
Device binding
Query accounts, device fingerprints and plan rights together.