I’m sharing BrowserWing, an open-source experiment exploring a more efficient way for AI agents to control browsers.

Most browser-based agents today rely on frequent LLM interactions for DOM inspection, screenshot reasoning, and step-by-step execution. While this works, it’s slow and consumes a large number of tokens.

BrowserWing takes a different approach. Browser automation is implemented once as scripts and then exposed as MCP commands. AI agents focus on expressing intent and simply call these commands, instead of interacting with the browser step by step.

This significantly reduces interaction latency and token usage, while keeping browser execution deterministic and controllable.

This is an early release, focused on validating the idea rather than feature completeness. I’m especially interested in feedback from people building AI agents, browser automation, or MCP-based tools.

Repo: https://github.com/browserwing/browserwing

Thanks for reading.