brw

brw provides a semantic browser control for agents, enabling them to interact with a real, visible Chrome instance. It exposes the browser through MCP (Multi-Agent Communication Protocol) and an HTTP JSON API, allowing various agent harnesses and custom clients to drive it. Agents operate using stable references (like e17) derived from DOM and accessibility data, rather than relying on brittle CSS selectors or re-interpreting screenshots.

The platform is designed for real, signed-in web work, bridging to an already authenticated Chrome profile. This means agents can access gated dashboards, signed-in applications, and content that sandboxed agent browsers cannot. After every action, brw returns a post-action observation, informing the agent of what actually happened, which leads to fewer turns, less token spend, and reduced wall time compared to pixel-based interpretation.

Key functionalities include semantic snapshotting, reading prose, links, headings, forms, tables, and structured product data directly from pages. It supports actions like clicking, typing, filling, selecting, scrolling, dragging, uploading, waiting, and asserting by stable references. brw also offers features like tab management, download control, console access, network capture and replay, and optional Set-of-Marks overlays on screenshots for visual fallback.