// Agent profile
Agent Browser is a native Rust CLI from Vercel Labs that gives AI agents a browser they can drive one shell command at a time. The agent calls snapshot to get the page as compact text with element refs like @e1, acts on a ref (click, type, scroll), then snapshots again to check the result. No Playwright or Node.js runtime is required for the daemon. It supports headless and headed Chrome, CDP attachment to running browsers, cloud browser providers (Browserbase, Browserless), and alternative engines like Lightpanda. At 47,860 GitHub stars, it is the most popular browser automation tool purpose-built for AI agents.
Agent Browser takes a different approach to browser automation for AI agents. Instead of taking screenshots and piping them through a vision model, it renders pages as compact accessibility-tree text where each interactive element gets a ref like @e1. The agent reads this text, decides what to do, and issues a single CLI command — agent-browser click @e1 or agent-browser type @e3 "hello" — then takes a new snapshot to see the result. This observe-act-observe loop runs through shell calls, making it compatible with any agent framework that can execute commands.
The CLI sends commands to a background Rust daemon that controls Chrome over the DevTools Protocol. Each session gets its own daemon and browser instance. The daemon starts on the first command and exits after an hour of inactivity, optionally saving session state. This architecture means agents pay zero startup cost after the first command and can maintain long-running browser sessions across many steps.
Snapshot-based page reading: The snapshot command returns the page as structured text with refs for every interactive element. The -i flag filters to only interactive elements. Refs describe the page at snapshot time — after navigation or DOM changes, agents take a fresh snapshot before acting.
Multiple browser engines: Chrome for Testing is the default, but agents can attach to a running Chrome via --cdp, use Lightpanda or Obscura via --engine, drive Mobile Safari in the iOS Simulator, or connect to cloud providers like Browserbase, Browserless, and AgentCore.
Agent skills system: The CLI ships versioned instruction files that agents can load with agent-browser skills get core. These teach the agent the command vocabulary, ref conventions, and common workflows. The skills update with the CLI version, so agents always have accurate instructions.
MCP server: Running agent-browser mcp starts a Model Context Protocol server over stdio, letting MCP-compatible clients (Claude Code, Cursor, etc.) use the browser through structured tool calls rather than raw shell commands.
Web scraping and data extraction: Agents navigate multi-page sites, fill search forms, paginate through results, and extract structured data — all through text-based refs rather than fragile CSS selectors or pixel coordinates.
End-to-end testing: QA agents can drive real browsers through user flows, checking page content through snapshots rather than screenshots, with deterministic element targeting via refs.
Browser-based workflow automation: Agents that need to interact with web applications — filing tickets, updating dashboards, managing cloud consoles — get reliable browser control without vision model overhead.
Rust-based desktop computer-use framework from Vercel Labs — agents control apps via accessibility trees instead of screenshots.
Builder.io's open-source framework for building agent-native applications — shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together.
Persistent memory layer for AI coding agents — benchmark-backed (95.2% on LongMemEval-S), 92% fewer tokens per session vs full-context pasting, zero manual memory.add() calls.