BrowserWing addresses the inefficiency of token-heavy LLM interactions for browser automation by enabling AI agents to directly control browsers via Model Context Protocol (MCP) commands. It targets developers and AI agents seeking faster, more reliable web automation, data extraction, and RPA tasks, offering a token-efficient alternative to traditional LLM-driven browser control.

How It Works

The system comprises a React/Vite frontend, a Go backend API, and a browser recorder/player. It captures user interactions (clicks, inputs, navigations) and converts them into MCP commands. This approach allows AI agents to execute browser actions directly, bypassing slow LLM interpretation loops. The architecture leverages a dedicated MCP server to interface with various LLMs (OpenAI, Claude, etc.) and maintains stable sessions through cookie and storage management, facilitating complex, replayable automation workflows.

Quick Start & Requirements

• Installation: Recommended: Download prebuilt binaries from Releases. Execute ./browserwing (Linux/macOS) or ./browserwing.exe (Windows) after chmod +x. Alternatively, build from source using make install and make build-embedded (requires Go and pnpm).
• Prerequisites: Google Chrome or Chromium installed and accessible.
• Access: Run http://localhost:8080 in a browser.
• Docs/Demo: browserwing.com (general website).

Highlighted Details

• Comprehensive browser management: launch, control, and persist sessions with cookies.
• Visual script recording: captures clicks, inputs, and navigations for editing and replay.
• Seamless MCP integration: converts recorded scripts into executable MCP commands for AI agents.
• Broad LLM support: configurable for OpenAI, Claude, DeepSeek, and others.
• Token-efficient design promising fast performance for complex web tasks.

Maintenance & Community

• Community: Active Discord server (https://discord.gg/BkqcApRj) and Twitter presence (https://x.com/chg80333).
• Contributions: Welcomes issues and PRs with clear reproduction steps or rationale; feature ideas are discussed via GitHub Discussions.

Licensing & Compatibility

• License: MIT License.
• Compatibility: The project includes a disclaimer stating it's intended for personal learning and legitimate automation, advising against illegal use or TOS violations, which may imply restrictions on certain commercial or aggressive scraping use cases.

Limitations & Caveats

The project's disclaimer explicitly prohibits illegal use or violation of site terms, framing its intended use for personal learning and legitimate automation. This suggests potential limitations for applications that might push the boundaries of website Terms of Service.