BrowserSkill: CLI+browser extension for AI agents

Tencent released BrowserSkill, a combination of CLI and browser extension that allows any AI agent with shell access to use the user's real, logged-in browser without interruption. It supports remote mode, session management, and a security model that prevents credential extraction.
Key insights
BrowserSkill lets AI agents use the user's real browser with login state
Unlike headless browsers that require separate test accounts, BrowserSkill operates on the user's daily browser profile with existing cookies and login states. Automation runs in a separate 'Agent Window' that does not interfere with the user's normal browsing. It is framework-agnostic and works with any agent that can execute shell commands.
Four-layer architecture with strict separation of concerns
The architecture consists of Agent → bsk CLI (shell) → bsk daemon (local IPC) → browser extension (WebSocket) → Agent Window. The agent never directly touches the browser. Sessions are the core abstraction: each session has a unique ID, dedicated Agent Window, session-level ref-store, and borrow table. Operations within a session are serialized for reference safety.
Security model prevents credential extraction and unauthorized access
Write operations are sandboxed to the Agent Window unless a tab is explicitly borrowed (with user confirmation). Two automation switches (confirm before borrow, allow help requests) are stored in the browser and cannot be overridden by CLI. File transfer splits authority: CLI handles paths, daemon manages storage, extension handles browser transactions. Credentials are never extracted.
Remote mode enables cloud agents to control local browser securely
Version 0.3.0 added remote mode: the agent runs on a server and connects to the user's local browser via a one-time pairing link with renewable device credentials over TLS-authenticated WSS. This allows cloud-based agents to use the user's local login state while maintaining security.