BrowserAct Skills: AI Browser Automation CLI

1. Project Overview

BrowserAct Skills is an open-source browser automation CLI that lets AI agents (Claude Code, Cursor, VS Code, and other shell-capable agents) drive a real browser to complete web tasks — including sites protected by anti-bot systems — and hand off to a human operator when the agent gets stuck.

2. Background & Positioning

Most browser-automation frameworks assume a clean, cooperative web: no CAPTCHAs, no fingerprinting, no login walls. In practice, AI agents that try to browse, search, or extract data from real-world sites constantly run into anti-bot defenses, geo-restrictions, and account-gated content that break naive automation. BrowserAct was built to close that gap — it treats "the agent will eventually get blocked" as the default case, not the exception, and designs around it.

Its core mission is to give agents a resilient, token-efficient interface to the live web, with a built-in escape hatch: when automation truly cannot proceed, control can be handed to a human via a live remote-control URL, and the agent resumes automatically once the human is done.

Compared with general-purpose browser-automation libraries (e.g. Playwright/Puppeteer wrappers), BrowserAct differs in three ways:

It ships a three-layer anti-bot strategy (environment, execution, human) rather than leaving evasion entirely to the caller.
It returns an indexed, text-based page state instead of raw DOM/HTML/JSON, which is far cheaper in LLM tokens and easier for agents to reason about.
It bundles a catalog of 30+ ready-made Skills for popular platforms (Amazon, YouTube, Google Maps, Instagram, etc.), so common scraping tasks don't require writing automation from scratch.

3. Feature Categories

🛡️ Anti-Bot Protection (3-Layer Approach) — mechanisms to keep automation running against defended sites.

Stealth browser fingerprinting to avoid bot-detection signatures
TLS fingerprint rotation
Proxy rotation for IP diversity
Automatic CAPTCHA solving
Live human-takeover URL when automation is fully blocked

Purpose: keep long-running or sensitive tasks alive even when a site actively resists automated access.

🌐 Browser Modes — different session identities depending on the task.

Chrome mode (reuses the local browser's existing login state)
Stealth privacy mode (fresh, unlinkable fingerprint per session)
Stealth fixed-identity mode (persistent fingerprint/account across sessions)

Purpose: match the automation identity to the task — logged-in workflows, anonymous scraping, or consistent long-term accounts.

🤖 Agent-Optimized Interaction — the core CLI primitives agents call directly.

Indexed interaction (click 3, input 5 "text") instead of raw selectors
Compact, token-efficient text state output
Multi-session, multi-browser concurrency without cross-task interference

Purpose: make browser control cheap and reliable for LLM-driven agents, not just human scripts.

📦 Solutions Catalog (30+ Pre-Built Skills) — ready-to-use automation for common targets.

Amazon, eBay, Etsy, Walmart (product/listing/review extraction)
YouTube, TikTok, Instagram, X/Twitter (content, comments, profile data)
Google Maps, LinkedIn, Indeed (leads, business info, job listings)
Reddit, Xiaohongshu, Douyin, Zhihu, WeChat (regional platform coverage)

Purpose: skip re-implementing scrapers for platforms the community already covers.

🛠️ Skill Forge — a companion tool that generates new Skills.

Explores a target site once, then produces a reusable, deployable scraping Skill

Purpose: turn one-off manual exploration into a repeatable, shareable automation asset.

4. Key Highlights

Human handoff without losing state — when an agent hits a wall (CAPTCHA, 2FA, manual verification), it can generate a live remote-control link; a human takes over from any device, and the agent resumes the session afterward instead of restarting.
Token-efficient page state — pages are represented as an indexed, text-formatted list of interactive elements rather than raw HTML/DOM, reducing LLM context cost significantly compared to typical automation output.
Three-layer anti-bot defense — stealth fingerprints, TLS/proxy rotation, and CAPTCHA solving are built into the CLI rather than left to the integrator to assemble.
Parallel, isolated sessions — multiple concurrent browser sessions and accounts run independently, useful for multi-account workflows or parallel data collection.
Cloud or local execution — tasks can run on BrowserAct's managed cloud infrastructure with no setup, or as local Skills integrated directly into an agent's toolset.
Mostly free to use — most commands require no signup or only a free login; only managed proxies and stealth browser sessions beyond a free tier are paid.

5. Use Cases by Role

General developers — build agent workflows or scripts that need to browse, click through, and extract data from real websites without hand-writing selectors for every anti-bot obstacle.

Data / research scientists — collect structured data (product listings, social posts, reviews, business info) from platforms like Amazon, Google Maps, or Instagram for analysis, using the pre-built Solutions Catalog instead of custom scrapers.

Project managers / growth & marketing teams — extract competitive intelligence, lead lists, or content trends (LinkedIn jobs, Product Hunt launches, Reddit discussions) without engineering a scraping pipeline from scratch.

6. Getting Started

Find what you need — browse the Solutions Catalog for an existing Skill covering your target platform, or check the docs for the full command reference:

# see docs/quick-start.md and the solutions/ directory in the repo

Install / integrate — tell your AI agent to install the Skill directly from the repository:

Install browser-act.
Skill source: https://github.com/browser-act/skills/tree/main/browser-act
Verify it works after installation.

Then run a first command:

browser-act stealth-extract https://example.com

Contribute — fork the repository, add or improve a Skill under solutions/, and open a pull request; new platform integrations and fixes to existing Skills are welcome.

7. Project Structure

skills/
├── browser-act/               # Core CLI implementation
├── browser-act-skill-forge/    # Tool for generating new scraping Skills
├── solutions/                  # Catalog of 30+ pre-built platform Skills
├── docs/                       # Full documentation (quick-start, command reference, etc.)
├── assets/readme/               # README media assets
├── .github/workflows/           # CI configuration
├── requirements.txt
└── LICENSE                      # MIT

browser-act/ contains the CLI agents actually call (browser-act <command>).
browser-act-skill-forge/ is the tool used to author new Skills for sites not yet in the catalog.
solutions/ is where ready-made, platform-specific Skills live — the fastest path to a working integration.

8. Related Ecosystem

Upstream/managed platform: api.browseract.com — the managed cloud service that powers hosted browser execution, proxies, and stealth sessions referenced by the CLI.
Agent hosts it integrates with: Claude Code, Cursor, VS Code, and other shell-execution-capable AI agents.
Complementary tools: general browser-automation libraries (e.g. Playwright, Puppeteer) address similar low-level browser control but without BrowserAct's anti-bot layer, agent-oriented indexed state, or pre-built Skills catalog.

9. License

MIT License.

✅ Free to use, copy, modify, merge, publish, and distribute, including in commercial and closed-source projects.
✅ Sublicensing and private modification are allowed.
❌ No warranty is provided; the authors are not liable for damages arising from use.
ℹ️ The original copyright and license notice must be retained in copies or substantial portions of the software.
ℹ️ The MIT license covers the CLI and Skills in this repository; BrowserAct's managed cloud service (proxies, stealth browsers beyond the free tier) is a separate paid offering, not covered by the open-source license.

10. FAQ

Q: Do I need to sign up to use BrowserAct?
A: Most core features work without signup or with only a free login. Paid usage is limited to managed proxies and stealth browser sessions beyond the included free tier.

Q: What happens when a site blocks the agent completely?
A: The CLI can generate a live remote-control URL so a human can take over the session from any device; once the human finishes, the agent resumes automatically.

Q: Do I have to write my own scraper for a specific website?
A: Check the solutions/ directory first — it already covers 30+ platforms (Amazon, YouTube, Google Maps, Instagram, LinkedIn, and more). For a new target, use Skill Forge to generate a reusable Skill from a single exploration session.

Q: Which AI agents can use this?
A: Any agent that can execute shell commands, including Claude Code, Cursor, and VS Code-based agents.

Q: Where do I find the full command reference?
A: See the docs/ directory in the repository, starting with docs/quick-start.md.

11. Quick Links

Repository: https://github.com/browser-act/skills
Documentation: https://docs.browseract.com
Quick Start guide: https://github.com/browser-act/skills/blob/main/docs/quick-start.md
Solutions Catalog: https://github.com/browser-act/skills/tree/main/solutions
Official website: https://browseract.com

12. Summary

BrowserAct Skills gives AI agents a practical, resilient way to operate real browsers against real websites — anti-bot defenses, human handoff, and token-efficient state included — instead of treating browser automation as a solved, cooperative problem. Developers building agent workflows, and data/research teams needing structured data from platforms like Amazon, YouTube, or Google Maps, can start from the 30+ pre-built Skills rather than writing scrapers from scratch, and use Skill Forge to extend coverage to new sites.