SkillsBot.org

Agent Browser Skill

Vercel Labs' browser automation CLI for AI agents. With this skill, Claude can open websites, click, fill forms, take screenshots, extract data and test web apps in Chrome.

Compared head to head: agent-browser vs Playwright CLI vs browser-use.

  • License Apache-2.0
  • Installs 1.5M
  • GitHub stars 51.6K
  • Audits 2 of 3 passed

Our hands-on test, by SkillsBot.org

We asked Claude to open a page, use its filter and take a screenshot

agent-browser worked end to end: Claude installed the CLI and its browser, loaded the skill's current usage guide, opened our All skills page, typed into the filter by element reference, read the result (7 of 416), took the screenshot and closed its session. Without the skill Claude reached the same answer by writing a Playwright script, slightly faster, because this machine already had Playwright installed. The skill matters most where no browser tooling is set up and for multi-step sessions.

  • InstallWorked2 s with the skills CLI
  • Picked up on its ownYesinvoked agent-browser
  • Time41 s31 s without the skill
  • API cost$0.20$0.11 without the skill

The task, word for word

Open https://skillsbot.org/all-skills in a browser, type 'pdf' into the filter box, take a screenshot as filter.png, and tell me how many skills match.

With the skill

Screenshot taken by agent-browser of the SkillsBot All skills page filtered by pdf, showing 7 of 416 skills

Without the skill

Screenshot taken with a Playwright script of the same filtered page
The screenshots each run produced. Same page state, same count. Full size: with the skill, without.

With the skill

Claude's answer with agent-browser listing all seven matching skills

Without the skill

Claude's answer without the skill
With the skill Claude read the accessibility snapshot and listed all seven matches, and explained why canvas-design matches “pdf”. Without it, Claude only described the first three visible in its screenshot. Full size: with the skill, without.

What it actually did

  1. Invoked agent-browser, found the CLI missing, ran npm i -g agent-browser and agent-browser install.
  2. Ran agent-browser skills get core to load the usage guide that matches the installed version, as the stub SKILL.md instructs.
  3. Opened the page in a named session, took an accessibility snapshot, filled the filter box by its @e reference and waited.
  4. Read the count from a fresh snapshot, saved filter.png and closed the session.

Where it helped

  • Element references from snapshots made the interaction exact, with no CSS selectors to guess.
  • Reading the accessibility tree gave Claude the full result list, not just what fitted on screen.
  • Cleaned up after itself by closing the browser session.

Limits we hit

  • Installs a global npm package and downloads a browser on first use, so the first run is slower and changes your machine.
  • The SKILL.md is only a stub; the real instructions come from the installed CLI, so you cannot review them before installing.
  • On a machine that already has Playwright, plain Claude did the same simple job faster and cheaper (31 s, $0.11 against 41 s, $0.20).

Use it for

  • Letting an agent check your web app after a change: click through, fill forms, screenshot.
  • Scraping data from pages that need clicks or a login you control.
  • Exploratory QA and bug hunts on a staging site.
  • Automating Electron desktop apps, which the skill also supports.

One run with the skill and one without, same prompt, fresh folders, Claude Code 2.1.325 in headless mode. How we test skills. Every skill we have tested.

Last updated . The agent-browser skill is made and maintained by Vercel; SkillsBot.org is an independent directory and is not affiliated with it.

Use agent-browser in Codex, Cursor and GitHub Copilot

agent-browser is a standard Agent Skills folder, so the same files work in other coding agents. Name the agent when you install:

Parts of a skill that rely on Claude Code features, such as subagents or / commands, may behave differently elsewhere. Paths for each agent are in our install guide.

Agent Browser Skill facts

Author
Vercel (@vercel-labs)
Repository
vercel-labs/agent-browser
License
Apache-2.0
Installs
1.5M
GitHub stars
51.6K
First seen on skills.sh
Jan 19, 2026
SKILL.md length
329 words
Other files
None

Security audits on skills.sh:

Installs, stars, first-seen date and audits from skills.sh on November 16, 2026. SKILL.md read from GitHub the same day.

What agent-browser does

agent-browser is a command-line tool that controls Chrome or Chromium through the DevTools protocol. Instead of raw HTML, it gives the agent an accessibility-tree snapshot with short element references such as @e5, which keeps pages compact enough to work with.

The skill triggers on requests like "open a website", "fill out a form", "take a screenshot", "scrape data from a page" or "test this web app". Its description also mentions automating Electron desktop apps such as VS Code, Slack, Discord, Figma, Notion and Spotify.

A stub that loads the real guide

The SKILL.md is short on purpose. It tells the agent to install the CLI and then load the usage guide that matches the installed version:

npm i -g agent-browser && agent-browser install
agent-browser skills get core

The core guide then lists further references for specialised tasks. This keeps the instructions in step with the CLI version you actually have.

Permissions

The skill declares allowed-tools: Bash(agent-browser:*), which pre-approves only agent-browser commands. Everything the agent does in the browser happens with your logged-in sessions if you use your own profile, so review what it is about to do.

Agent Browser Skill: FAQ

How do I install agent-browser?

Run npx skills add vercel-labs/agent-browser --skill agent-browser for the skill, then npm i -g agent-browser && agent-browser install for the CLI and browser.

Is agent-browser free?

Yes. It is open source under the Apache 2.0 license.

How is agent-browser different from Playwright?

Playwright is a general browser automation library. agent-browser is a CLI designed for agents, returning compact snapshots with element references that are easier for a model to act on.