AgentSkillsDir home
Submit
agent-browser project image
Claude skills

Install agent-browser CLI + skill

agent-browser from Vercel Labs (vercel-labs/agent-browser, Apache-2.0) is a browser automation CLI for AI agents: a fast native Rust binary that drives Chrome over CDP and hands your agent accessibility-tree snapshots with compact @eN element refs. A matching discovery skill tells Claude Code, Codex or Cursor when to reach for it.

This is a source-linked directory entry, not Vercel’s official site, not a Playwright MCP ranking and not a hosted browser cloud. AgentSkillsDir does not run browsers or automation for you.

  1. Install the CLI and Chrome

    Global npm install of the native binary; agent-browser install downloads Chrome for Testing on first run.

    npm i -g agent-browser && agent-browser install
  2. Add the discovery skill

    Fetched from the repository, so it stays current. Do not copy SKILL.md out of node_modules.

    npx skills add vercel-labs/agent-browser
  3. Then your agent loads the real guide

    The installed skill is a stub; it points the agent at this command.

    agent-browser skills get core
  • npx skills add vercel-labs/agent-browser — fetched from the repo, stays current
  • node_modules/…/SKILL.md — copying it by hand goes stale
vs Playwright MCP?

Upstream README and SKILL.md commands, not yet run by AgentSkillsDir.

What it is, and what it is not

  • It is a CLI your agent calls from the shell: open a page, take a snapshot, click or fill by ref, screenshot, close. Upstream also lists sessions, an authentication vault, state persistence and video recording.
  • It is a skill in the Agent Skills format, but a thin one: the repo SKILL.md only exists so agents discover the CLI. The real instructions come from the CLI itself.
  • It is not a Playwright MCP server ranking, not a hosted browser service you buy here, and not a crawler. Use it on sites you own or are allowed to automate.

Step 1: install the CLI

Commands follow the upstream README and SKILL.md at commit 5868f7c05020 (fetched 31 Oct 2026). AgentSkillsDir has not run them yet; treat every step as upstream documentation.

npm i -g agent-browser && agent-browser install

npm i -g installs the native binary. agent-browser install then downloads Chrome from Chrome for Testing (first time only); existing Chrome, Brave, Playwright and Puppeteer browsers are detected automatically. On macOS below 13 it stops and asks you to update or install Chrome manually.

Other routes from the README, if global npm is not for you:

  • Project dependency (pin the version)

    npm install agent-browser agent-browser install

  • Homebrew (macOS)

    brew install agent-browser agent-browser install

  • Cargo (Rust)

    cargo install agent-browser agent-browser install

  • Linux system libraries

    agent-browser install --with-deps

    Uses your package manager (sudo when not root) and exits nonzero if a browser library cannot be installed.

  • Upgrade later

    agent-browser upgrade

    Detects npm, Homebrew or Cargo and runs the matching update.

Step 2: add the skill

npx skills add vercel-labs/agent-browser

The README says this works with Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode, Windsurf. In Claude Code it writes a stub to .claude/skills/agent-browser/SKILL.md. AgentSkillsDir has not checked each assistant yet.

Do not copy SKILL.md from node_modules. Upstream says that copy will go stale; npx skills add fetches the skill from the repository instead.

Why the skill is a stub: skills get core

A SKILL.md file cannot change between CLI releases, so upstream keeps it minimal and lets the CLI serve the guides bundled with the installed version. Before its first browser command, the agent runs:

agent-browser skills get core          # start here
agent-browser skills get core --full   # + full command reference
agent-browser skills list              # what this CLI version ships

Specialized guides named in the stub (AgentSkillsDir has not checked these on the current CLI):

core
Start here: workflows, common patterns, troubleshooting. Add --full for the command reference and templates.
electron
Electron desktop apps such as VS Code, Slack, Discord or Figma.
slack
Slack workspace automation.
dogfood
Exploratory testing, QA and bug hunts on your own app.
derive-client
Record a HAR and derive a standalone API client for a site.
vercel-sandbox
Run agent-browser inside Vercel Sandbox microVMs.
protected-vercel-deployments
Access protected Vercel deployments.
agentcore
AWS Bedrock AgentCore cloud browsers.

Minimal smoke test

From the README Quick Start; AgentSkillsDir has not run it yet.

agent-browser open example.com
agent-browser snapshot   # accessibility tree with @eN refs
agent-browser close

The snapshot prints the accessibility tree with refs such as @e2; the agent then runs click @e2 or fill @e3 "…" against those refs, and takes a fresh snapshot after the page changes.

agent-browser vs Playwright MCP

They reach the browser by different routes. agent-browser is a native CLI plus a skill: the agent runs shell commands and reads compact snapshots, and the daemon needs no Playwright, Puppeteer or Node.js. Playwright MCP is Microsoft’s MCP server, which the client connects to as a set of browser tools. agent-browser can also expose an MCP server (agent-browser mcp), but this page covers the CLI and skill path. Pick agent-browser if you want the Vercel Labs CLI workflow in your shell; if you are choosing an MCP server, see our Playwright MCP entry.

Before you install

  • This is third-party code that controls a real browser, including logged-in sessions and a credential vault. Read the source and the core guide before giving an agent access.
  • Only automate sites you own or are permitted to automate, and respect their terms.
  • Listing is not a security review. AgentSkillsDir has not reviewed this code.

FAQ

What is the agent-browser skill?

The discovery skill that ships with Vercel Labs’ agent-browser CLI (vercel-labs/agent-browser, Apache-2.0). It tells your coding agent when to drive a real browser through the CLI, which controls Chrome over CDP and reads pages as accessibility-tree snapshots. It is not a Playwright MCP leaderboard.

How do I install the CLI and the skill?

Run npm i -g agent-browser && agent-browser install, then npx skills add vercel-labs/agent-browser. Taken from the upstream README on 31 Oct 2026; not yet run by AgentSkillsDir.

Why is the installed SKILL.md so short?

It is a deliberate discovery stub. It only points the agent at agent-browser skills get core, which prints guides bundled with the installed CLI, so the instructions always match your version.

Can I copy SKILL.md from node_modules instead?

No. Upstream warns that a copy from node_modules goes stale. Use npx skills add for the stub and agent-browser skills get for the content.

agent-browser or Playwright MCP?

Different paths. agent-browser is a native CLI plus a skill: your agent runs shell commands, with no Playwright or Puppeteer dependency. Playwright MCP is Microsoft’s MCP server that exposes browser tools to the client. Our /mcp/playwright-mcp entry covers that one; there is no separate comparison page.

Do I need Node.js to run the browser daemon?

Upstream says no: the daemon is a native Rust binary and needs no Playwright or Node.js. npm is just one way to install it (Homebrew and Cargo also work); Node.js 24+ is only needed to build from source.

Which assistants work with npx skills add?

The README lists Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot, Goose, OpenCode and Windsurf. AgentSkillsDir has not checked each one yet.

Does AgentSkillsDir host browsers or sell cloud automation?

No. This is a source-linked directory entry. Nothing installs or runs from this page. Upstream documents Vercel Sandbox and AWS AgentCore options separately.

Has AgentSkillsDir reviewed the code?

No. Listing is not a security review. Read the source before installing, and only automate sites you own or are allowed to use this way.

Are the star and version numbers current?

They are a dated snapshot from the GitHub API and the npm registry; see the last-checked time on the source card.

Corrections and suggestions

Spotted something out of date, or want to suggest a skill? Email a correction. To suggest a resource, submit it for manual review.