Agent Browser Protocol (ABP) is a Chromium fork with MCP and REST APIs embedded directly in the browser engine, designed to bridge the gap between continuous async web browsing and the step-by-step reasoning of LLM agents. Each HTTP request represents one atomic action: ABP injects native input, waits for the page to settle, captures a compositor screenshot with virtual cursor, collects events (dialogs, file choosers, downloads), then freezes JavaScript and virtual time until the next step. This eliminates the need for WebSocket session management, manual waits, and polling. It achieves 90.53% on the Online Mind2Web benchmark. Key features include engine-level JS execution pause, virtual time control, automatic per-action screenshots with element markup, session recording to SQLite for agent fine-tuning, and an embedded MCP server with 18 tools. Compatible with Claude Code, Codex CLI, Opencode, and any MCP client.

12m read timeFrom github.com
Post cover image
Table of contents
What you get per actionQuick StartABP in ActionWhy ABP (and why a Chromium build)What Makes ABP DifferentComparisonCommand Line FlagsProject StructureStatusTestingREST APIMaintainersUninstallContributingLicenseAcknowledgments

Sort: