Agent Browser Protocol (ABP) is a Chromium fork that takes an unusual approach to browser automation: rather than layering a control framework on top of an existing browser, developer 'theredsix' has embedded both an MCP server and an HTTP REST server directly into the browser engine.
The practical difference matters. Tools like Playwright and Puppeteer were designed for test automation — scenarios where a human can intervene if something breaks. When used in LLM agent workflows, the asynchronous nature of a live browser becomes a reliability problem: the page keeps updating while the model is deciding what to do next. ABP sidesteps this by freezing JavaScript and virtual time between actions. Each API call injects native input events, waits for the page to fully settle, captures a compositor screenshot with cursor position, collects browser events, and holds everything still until the next instruction arrives. The overhead is around 100ms per action.
The benchmark results are notable. ABP scored 90.53% on Online Mind2Web, a real-world browser task evaluation from OSU's NLP Group — up from the previous public leaderboard leader of 78.7%. Running on Claude Opus 4.6, it hit 96.15% on easy tasks, 89.86% on medium, and 85.51% on hard, filtering out tasks that human evaluators marked as impossible. The full test harness, including model selection, task filtering, and per-task logging, is published in a companion repository, 'abp-online-mind2web-results', for anyone who wants to verify the numbers independently.
Setup is straightforward. Adding ABP to Claude Code takes a single terminal command — 'claude mcp add browser -- npx -y agent-browser-protocol --mcp' — after which the browser is available as 13 native tools: navigate, click, type, screenshot, and others. Codex CLI and Opencode are supported out of the box. The MCP server is reachable by any MCP client via streamable HTTP at localhost:8222/mcp, and a REST API is available for teams that prefer to skip MCP entirely. The npm package handles binary download and launch automatically.
MCP has become a de facto interoperability layer across the major AI coding tools, which gives ABP's architecture some reach. With the server embedded at the engine level rather than bolted on as an afterthought, it offers a cleaner foundation for agent workflows that depend on consistent, reproducible browser state.