dondai1234/agent-browser

agent-browser is a browser automation tool built for AI agents, not for humans to operate directly. It runs as a single Go binary with no Node.js, Playwright, or Puppeteer required, and it connects to AI assistants through the Model Context Protocol (MCP). When an AI needs to browse the web, fill a form, or pull data from a page, it calls this tool, which carries out the action and reports back a compact summary of what changed. The main design goal is reducing how many tokens the AI spends interpreting the browser. Other browser MCP servers send verbose accessibility trees after every action, which can cost tens of thousands of tokens per page. agent-browser sends compact reference lines and only what changed since the last step. On a measured login task, it used about 154 tokens compared to roughly 1,500 for Playwright MCP. It also exposes only 9 tools versus 22 or more for competing options, so connecting costs fewer tokens too. The tool gives the AI nine actions: act handles clicking, filling, and selecting any control by describing it in plain words, js runs JavaScript on the page and returns structured data as JSON rather than a list of page references, login handles a full sign-in sequence including multi-step flows like those used by Google and Microsoft, see takes a page snapshot, nav goes to a URL, and scroll, mark, and history round out the set. The login tool checks the resulting page state rather than just whether a click went through, so it distinguishes a successful login from a wrong-password error or a two-factor prompt. Version 3.2 added automatic cookie banner dismissal on every page load, support for a type of custom dropdown that standard accessibility tools cannot reach, and corrections to browser signals that can reveal a headless browser to bot-detection systems. The tool is licensed under MIT and installs as a single binary on macOS, Windows, and Linux. Configuration involves adding it to an MCP config file.