U Skills & Agents
← All Skills
Coding v5185339 By Vercel Labs

Agent Browser

A third-party skill from Vercel Labs that drives a native Rust browser-automation CLI (agent-browser). Claude calls it through Bash to interact with real websites and Electron desktop apps over the Chrome DevTools Protocol: navigating pages, filling forms, clicking buttons, taking screenshots and PDFs, intercepting network traffic, managing cookies and auth state, and handling multiple tabs. Instead of raw DOM scraping it uses compact accessibility-tree snapshots with stable element references, which makes interactions more reliable. The skill includes a core command set plus specialized add-ons for Electron apps (VS Code, Slack, Discord, Figma), Slack workspaces, QA/testing, and cloud browsers (Browserless, Browserbase, Vercel Sandbox, AWS Bedrock AgentCore). Use it when the user needs to automate a browser, test a web app, scrape or extract data, or reproduce a UI flow. Requires Node.js 24 or newer and installing the pinned agent-browser CLI (npm i -g [email protected], then agent-browser install, adding --with-deps on Linux), plus Bash access to it. Note the pinned SKILL.md is a discovery stub: the detailed workflows are served by the CLI at runtime via agent-browser skills get core and track the installed version, which is why the npm package is pinned to the reviewed release. Licensed Apache-2.0 by Vercel.

browserautomationtestingcliscrapingweb
Install

Install for Claude Code (local)

Local Claude Code workflow only. Needs Node.js 24 or newer (the package declares engines.node >=24.0.0). On Linux, also install the browser's system libraries with the --with-deps flag shown below.

This is a Claude Code (local terminal) skill: it drives the host-installed agent-browser CLI through Bash. It is not meant for the claude.ai or Claude Desktop skill surfaces, where skills run in an isolated code-execution container that can't reach your host binary or control your local browser. The pinned SKILL.md below (commit 5185339) is a discovery stub: the actual workflows and command reference are served by the CLI itself via 'agent-browser skills get core' and always match the installed CLI version. Because of that, pin the npm package to the reviewed version (0.27.2) so a later release can't silently change the served instructions. Requires Node.js 24+. Licensed Apache-2.0 by Vercel; keep the LICENSE with the file if you redistribute it.

  1. 1

    Check Node.js

    Confirm Node.js 24+ with node -v. The CLI declares engines.node >=24.0.0 and will error or misbehave on Node 20/22.

    node -v   # expect v24.0.0 or newer
  2. 2

    Install the CLI (pinned version)

    Install the reviewed version and a browser. On Linux add --with-deps so the first browser launch has the system libraries it needs; on macOS/Windows drop that flag.

    npm i -g [email protected]
agent-browser install            # macOS / Windows
agent-browser install --with-deps  # Linux: also install system deps
  3. 3

    Add the skill to Claude Code

    Save the pinned SKILL.md to .claude/skills/agent-browser/SKILL.md in your project (or ~/.claude/skills/ for all projects). The deeper workflows load at runtime from the installed CLI via agent-browser skills get core, so keep the CLI pinned to the version you reviewed.

  4. 4

    Allow the tool

    Grant Bash access to the globally installed binary with Bash(agent-browser:*), then ask Claude to automate a browser task. Don't allow an unpinned npx agent-browser form: in a project without it as a local dependency, npx would resolve and run the latest unreviewed release, defeating the version pin. If you must use npx, pin it: npx [email protected].

View the repo
Toolkit

What's in this toolkit

This is a toolkit of multiple skills, each with its own SKILL.md. Open the source on GitHub (pinned to the reviewed commit) to read the instructions for the skills you install.

Agent Browser

A third-party skill from Vercel Labs. The pinned SKILL.md is the source of truth; this entry is a short summary.

It lets Claude drive the agent-browser Rust CLI through Bash to automate real browsers and Electron apps over the Chrome DevTools Protocol: navigation, form fills, clicks, screenshots, PDFs, network interception, cookies, auth state, and multi-tab work, using accessibility-tree snapshots with stable element refs for reliable interactions. Specialized command sets cover Electron apps, Slack, QA/testing, and cloud browsers.

Requires Node.js 24+ and the pinned CLI (npm i -g [email protected], then agent-browser install, adding --with-deps on Linux) plus Bash access to the binary. The pinned SKILL.md is a discovery stub; the real workflows are served by the CLI at runtime via agent-browser skills get core and match the installed version, so keep the package pinned to the reviewed release (0.27.2). Licensed Apache-2.0 by Vercel.