Tencent Releases BrowserSkill: Open-Source Tool for AI Agents to Control Logged-In Browsers

Tencent has open-sourced BrowserSkill, a tool designed to allow AI agents to interact with real, logged-in browser sessions rather than isolated sandbox environments. The project, which has already garnered 7.5k GitHub stars, integrates with shell-capable agents such as Cursor, Claude Code, Codex, and DeepSeek Harness through a command-line interface (bsk) and a browser extension.

Core Functionality and Architecture

BrowserSkill operates by launching a visible Agent Window where the AI performs tasks. It can borrow existing tabs from the user's Chrome or Microsoft Edge browser, execute actions like form filling, document reading, or internal site searching, and then return the tabs to their original state. This approach allows agents to leverage the user's existing authentication and session state, bypassing the need for complex credential management or headless browser configurations.

Key capabilities include:

• Browser Automation: Clicking, typing, tab management, and full-page screenshots.
• Web Debugging: Agents can inspect network requests, console messages, and performance data, with support for request replay and response modification.
• Human-in-the-Loop: The system supports prompts for human intervention during login or verification steps, ensuring security and control.
• Remote Execution: Supports pairing local browsers with server-side agents via authenticated WebSockets, allowing agents to run on servers while controlling local browsers.

Integration and Security

The tool provides specific plugins for DeepSeek Harness (DSH), offering native browser_* tools and task previews. For other agents, integration is handled via the bsk CLI, which manages the background daemon and skill installation.

Security and privacy are addressed through several mechanisms:

• No Mandatory Cloud: The tool does not operate a mandatory cloud service or collect product telemetry.
• Audit Logs: Operation audits are stored locally in BSK_HOME/audit (disabled by default) and expire after 30 days.
• Data Retention: Debugging evidence is stored in the browser profile with a retention budget of 50 records / 50 MiB, evicting older records when the limit is reached.
• Permission Boundaries: Agents operate with the permissions of the signed-in website, requiring users to trust the agent's tasks. Settings allow users to confirm before the agent borrows tabs or requests human help.

Technical Requirements

BrowserSkill supports macOS (Apple Silicon, Intel), Linux (x64, ARM64), and Windows (x64). The browser extension requires Chromium 125 or higher. The project is built with Rust and pnpm, and is released under the MIT License.