BrowserSkill: Open-Source Tool for AI Agents to Use Your Real Browser

Tencent open-sourced BrowserSkill, a combination of CLI and browser extension that lets any AI agent capable of shell commands use the user's real, logged-in browser without interruption. It runs automation in a separate Agent Window, reuses login state, and supports human intervention for CAPTCHAs. The architecture includes a daemon, WebSocket connection, and strict security model.
Key insights
Architecture: Four-Layer Chain with Strict Separation
BrowserSkill's architecture consists of Agent → bsk CLI (shell) → bsk daemon (local IPC) → browser extension (WebSocket) → Agent Window. The agent never directly touches the browser. The daemon manages session routing and serializes RPC per session. The extension uses CDP to drive the browser, managing the Agent Window, element references, and tab borrowing. Sessions are the core abstraction: each has a dedicated window and ref-store.
Security Model: Permission Convergence and Credential Protection
BrowserSkill implements multiple security layers: all write operations are sandboxed to the Agent Window unless tabs are explicitly borrowed with user confirmation. Two automation switches (borrow confirmation, help request) are stored in the browser and cannot be overridden by CLI. File transfer splits permissions among CLI, daemon, and extension. The skill explicitly states it never extracts credentials, cookies, or tokens. Remote mode uses one-time pairing and revocable device credentials.
Agent Engineering: Defensive Rules in SKILL.md
The SKILL.md file contains defensive rules for agents: navigation invalidates references, so agents must re-observe before interacting; borrowing refusals must be respected; help requests cannot be bypassed; screenshot evidence and page assertions are verified separately. Canvas interactions use single-use capture IDs with 2-minute expiry. The installation is agent-friendly: a single command triggers CLI install, skill injection, diagnosis, and verification.