Browser Automation and Web Testing Skills

The most installed browser skill in the registry, and for good reason. agent-browser is a CLI that talks to Chrome over the DevTools Protocol directly, with no Playwright or Puppeteer layer in between. Its defining idea is the snapshot-and-ref workflow: instead of feeding raw HTML to the model, it renders the page as a compact accessibility tree where every interactive element gets a short ref like @e3. A typical page view costs the agent roughly 200 to 400 tokens.

The core loop is four commands: open a URL, snapshot the page, act on a ref, snapshot again. Refs go stale the moment the page changes, and the skill is explicit about re-snapshotting after every navigation or dynamic re-render, which is exactly the kind of guidance that keeps agents from flailing. Beyond the core, companion sub-skills cover Electron apps, Slack automation, Vercel Sandbox microVMs, and AWS Bedrock AgentCore cloud browsers.

• Accessibility tree snapshots with @eN refs keep page context tiny
• Parallel browser sessions for multi-agent work
• Sub-skills for Electron, Slack, sandboxes, and cloud browsers
• Installs with one command, no browser framework dependency

The official Playwright team's answer to agent-driven browsing. playwright-cli wraps the full Playwright engine in small single-purpose commands: open, goto, click, fill, drag, press. Like agent-browser it works from page snapshots with element refs, so the agent reads a condensed view of the page rather than raw markup.

Because it is real Playwright underneath, you inherit the engine's strengths: reliable auto-waiting, cross-browser support, and first-class form handling, including a fill command with a submit flag that presses Enter for you. If your project already runs Playwright tests, this skill lets the agent reuse the same runtime it will be writing tests for.

• Maintained by the Playwright team at Microsoft
• Snapshot refs plus auto-waiting from the Playwright engine
• Covers clicks, double clicks, drag and drop, and form fills
• Natural companion to Playwright test suites

browser-use takes a scripting-first approach. The agent writes short Python snippets in heredocs and pipes them to a daemon that controls Chrome over CDP. Helper functions are pre-imported, so a working command can be as small as printing page_info() after opening a tab.

Its standout trait is attaching to your already running Chrome, existing profile and logins included, which makes authenticated flows much less painful. A doctor command diagnoses connection problems, and an opt-in library of site-specific domain skills teaches the agent proven interaction patterns for particular websites before it starts improvising.

• Attaches to your running Chrome with real sessions and cookies
• Python heredoc interface with pre-imported helpers
• Opt-in domain skills encode per-site automation knowledge
• Built-in diagnostics via a doctor command

browser-act runs a complete browser engine with an unusual focus on safety and privacy. Its Confirmation Gate protocol requires explicit user approval before creating or deleting browsers and before sensitive operations like logins, form submissions, and file uploads. All cookies, credentials, and page content stay local; the only thing that ever leaves your machine is a challenge image if you invoke its verification assistance.

It shines on the messy parts of real-world automation: isolated multi-browser profiles for parallel multi-account work, XHR and HAR network capture, content that only loads on scroll or click, and human-in-the-loop handoffs through headed mode when a step genuinely needs a person. It also offers a lightweight extraction mode that fetches JavaScript-rendered content without opening a full session.

• Confirmation Gate asks before any sensitive operation
• Local-only storage of sessions, cookies, and credentials
• Multi-browser isolation for parallel multi-account flows
• Network capture with XHR, fetch, and HAR output

Browserbase's skill drives the browse CLI, which can run a clean local browser, attach to a debuggable Chrome you already have open, or spin up a remote session in Browserbase's cloud. The environment is chosen per command, so development against localhost and production runs against hardened remote browsers can share one workflow.

The remote mode is the differentiator. Cloud sessions come with automatic CAPTCHA solving, residential proxies, and Browserbase Identity for verified browsing, which is what you want when a target site actively resists automation. Install counts are still small compared to the giants above, but the capability set is unique in this list.

• Switch between local Chrome and remote cloud sessions per command
• Automatic CAPTCHA solving and residential proxies in remote mode
• Attach to an already running debuggable Chrome
• Good fit for JavaScript-heavy and bot-protected sites

Anthropic's official toolkit for testing local web applications. Rather than exposing a fixed command set, it has the agent write native Python Playwright scripts, guided by a decision tree: read static HTML directly when possible, and fall back to a reconnaissance-then-action pattern for dynamic apps, navigating first, inspecting the rendered DOM, then acting on discovered selectors.

A bundled with_server.py helper manages server lifecycle around test runs, including multiple servers at once, so the agent can boot your dev stack, run assertions against it, and tear it down cleanly. The skill also encourages screenshot capture and browser log inspection, which is exactly the evidence an agent needs to verify frontend behavior rather than assume it.

• Official Anthropic skill, tuned for how Claude tests UIs
• Server lifecycle helper boots and stops your dev stack
• Reconnaissance-then-action pattern for dynamic apps
• Screenshots and browser logs as verification evidence

This one does not drive a browser at all. It is a knowledge skill from the team at Currents that loads the right reference guide for whatever Playwright work the agent is doing, organized as an activity-based index across more than forty focused documents: locators, fixtures, Page Object Model, flaky test debugging, CI configuration, API mocking, authentication flows, and more.

The coverage goes deep into territory most test suites eventually hit: accessibility testing with axe-core, visual regression, canvas and WebGL, service workers, WebSockets, clock mocking for date-dependent features, security testing for XSS and CSRF, performance budgets with Web Vitals, even Electron apps and browser extensions. Paired with any automation skill above, it is the difference between tests that pass once and suites that hold up in CI.

• Activity-based index across 40 plus reference guides
• Deep coverage of flaky tests, fixtures, and Page Object Model
• Accessibility, visual regression, and security testing included
• Written by Currents, a Playwright CI observability company

The Chrome DevTools team's own MCP server, with a skill that teaches agents to use it well. It exposes the debugging surface the DevTools panel gives humans: page snapshots with stable element uids, network request inspection, performance tracing, and script evaluation, with opt-in flags for extension tooling and memory debugging.

The skill encodes real workflow discipline: navigate, wait, snapshot, then interact, and route large outputs like traces and screenshots to files instead of flooding the context window. Sibling skills in the same repo go further, covering accessibility debugging, LCP optimization, and memory leak hunting. Install counts are modest, but this is the most authoritative option for debugging work.

• Built and maintained by the Chrome DevTools team
• Performance traces, network waterfalls, and memory tooling
• File output options keep large traces out of agent context
• Sibling skills for a11y debugging and LCP optimization

Firecrawl's CLI skill gives agents real-time web search, page scraping, whole-site crawling, and interactive fetching for pages that need clicks or logins, all returning clean markdown sized for LLM context windows. It runs against Firecrawl's hosted infrastructure, so the agent does not manage browsers at all.

It installs as a family: the core CLI skill plus focused sub-skills for scrape, search, crawl, map, download, and an agent mode, alongside build skills for integrating Firecrawl into application code and workflow skills for outcomes like deep research and SEO audits. A keyless free tier works out of the box, and an API key raises the limits and concurrency. Summed across its sub-skills this is one of the most installed tool families in the whole registry.

• Search, scrape, crawl, and interact from one CLI
• Clean markdown output tuned for LLM context windows
• Keyless free tier, API key raises limits
• Sub-skills for integration work and outcome workflows

just-scrape is ScrapeGraphAI's CLI skill, and its quarter-million installs make it the most adopted single scraping skill in the registry. Beyond search, scrape, and crawl, its signature move is schema-driven extraction: describe the JSON you want and the service uses AI to pull matching structured data out of the page, which spares the agent from writing brittle selectors.

It also covers a niche most scraping tools skip: monitoring pages for changes over time, plus request history inspection and credit management from the same CLI. It requires a ScrapeGraph API key, and the skill walks the agent through validating the setup with a small request before doing real work.

• AI-powered extraction of structured JSON from any page
• Page change monitoring built into the same CLI
• Request history and credit inspection commands
• Setup validation flow before real workloads