Labsco
Labsco/Repos/vercel-labs/agent-browser
REPO PACKAGE

vercel-labs agent-browser

COMMUNITY
vercel-labs · publisher37,903 repository starsApache-2.0 · Freegithub.com/vercel-labs/agent-browser
7skills
0ready to use

The core skill documents agent-browser's snapshot-and-ref workflow — open a page, take an accessibility-tree snapshot, then click, fill, or read elements by the refs it returns — plus screenshots, PDF export, batch execution, and a natural-language chat mode. The other six skills are each a guide to running that same CLI in one specific place: inside Electron desktop apps (VS Code, Slack, Discord, Figma, Notion, Spotify) over Chrome DevTools Protocol, inside a Vercel Sandbox microVM, against a Slack workspace, or on AWS Bedrock AgentCore's cloud browsers. dogfood is the one skill that is a workflow rather than a place: point it at your own app and it produces a bug report with reproduction screenshots and video.

This suits a developer or agent that has already decided to use agent-browser and needs to know which of its surfaces applies to the job at hand — VS Code automation, a Vercel deployment, an AWS-hosted session, or a Slack task. It is not a general browser-automation primer; the assumption throughout is that agent-browser itself is the tool, installed via npm, Homebrew, or Cargo, with its own bundled Chrome for Testing binary rather than your system Chrome.

READ THE FULL ANALYSIS

What it costs you to start. Five of the seven skills need nothing beyond the CLI's own installer, which downloads that private Chrome build the first time you run agent-browser install. The other two want more: AgentCore needs AWS credentials for its managed cloud browsers, and the Vercel Sandbox path needs a Vercel account to run the microVM. Linux users also need the installer's --with-deps flag for the browser's system libraries, which it checks for and fails on loudly rather than silently.

One structural note. Six of the seven skills exist because agent-browser runs in six different places, not because they teach six different tasks — read core first regardless of which environment you land on, since it is the only one that explains the ref-based interaction model the other six assume you already know.

37,903Stars on the GitHub repository, at last check.