BrowserAct: Open-Source Browser Automation for AI Agents

BrowserAct is a browser automation CLI built to serve AI agents on real-world web tasks where standard tools fail. The core challenge it addresses is that websites deploy anti-bot protections, agents get stuck, and handoff to humans is cumbersome. BrowserAct solves this through four design pillars.

First, it breaks through anti-bot blocks using three progressive layers. The environment layer applies stealth fingerprint spoofing, TLS rotation, and proxy switching—the tool notes that the vast majority of blocks never trigger with these measures alone. The execution layer provides solve-captcha (auto-solving CAPTCHAs) and stealth-extract (pulling protected pages in one command). The human layer generates a live URL for remote-assist, allowing a user to take over from any device while the agent resumes seamlessly when done.

Second, BrowserAct offers three browser modes by real-world scenario. Chrome mode reuses local Chrome login state via profile import or Chrome DevTools Protocol (CDP) attach. Stealth privacy mode enables frictionless batch scraping without login by issuing a fresh fingerprint per session plus proxy rotation with zero residue. Stealth fixed identity supports logged-in accounts and multi-browser parallel work by maintaining a stable fingerprint, stable IP, and stable account identity without bot flags.

Third, it implements zero-interference concurrency: cross-browser parallel isolation (independent cookies, fingerprints, proxies so sites cannot correlate them), same-browser multi-session (shared login state with independent execution so tasks do not block each other), and privacy mode (fresh fingerprint and empty profile per session). Fourth, the design is tailored for agent reasoning, not human scripts. It outputs compact indexed text (several times more token-efficient than JSON or HTML), returns state as an indexed list so agents click or input by number without DOM parsing, attaches semantic memory to each browser for task matching, and ensures concurrency safety through session ownership and explicit naming.

Installation is straightforward: users tell their AI agent to install browser-act from https://github.com/browser-act/skills/tree/main/browser-act and verify it works. The agent then invokes commands like `browser-act stealth-extract https://example.com` for protected-page extraction or `browser-act --session my-task browser open <id> https://example.com` followed by `browser-act --session my-task state` to see clickable elements, then `browser-act --session my-task click 3` to click by index. At the start of each session, agents run `browser-act get-skills core --skill-version 2.0.2` to discover environment state, browser list, and available commands.

BrowserAct is compatible with Windows, macOS, and Linux, and works with Claude Code, Cursor, VS Code, OpenCode, OpenClaw, Codex, Gemini CLI—any agent that can execute shell commands and load Skills. Pricing is almost entirely free: basic browser automation (Chrome/Chrome-direct), stealth extraction, CAPTCHA solving, remote-assist, privacy mode, and Skill Forge all require no signup. Login-only free features unlock stealth browsers (up to 5), solve-captcha, remote-assist, privacy mode, and Skill Forge. Paid features add stealth browsers beyond 5 and managed proxies (Dynamic/Static). Security is enforced through confirmation gating: sensitive operations (browser create/delete, profile import, proxy changes, security and privacy toggles) require explicit user approval each time, and prior approvals do not carry over.

A companion tool, Skill Forge, addresses the data-extraction bottleneck. Instead of writing scrapers by hand for repeated tasks on the same website, Skill Forge explores a site once, discovers its APIs and data patterns, generates a deploy-ready Skill package, and then runs reliably without re-exploration—handling 500 or 5,000 records through the same stable path. Users install browser-act-skill-forge from https://github.com/browser-act/skills/tree/main/browser-act-skill-forge and tell their agent "Forge a Skill that extracts job listings from LinkedIn — title, company, salary, URL. I'll run 300 keywords later." A Solutions Catalog provides 30+ pre-built Skills already generated by Skill Forge for sites including Amazon, Google Maps, YouTube, Reddit, WeChat, and Zhihu. Users can also generate custom Skills for any website in minutes with no coding required.

BrowserAct addresses a practical friction point in AI agent deployment: most websites employ anti-bot defenses that block standard automation tools, and agents often lack a graceful way to escalate to human operators when they encounter obstacles. The tool's architecture reflects this reality by layering three distinct blocking-defeat mechanisms—environment (fingerprint spoofing and proxy rotation), execution (CAPTCHA solving and protected-page extraction), and human (live handoff URLs)—rather than betting on a single approach.

The design philosophy shifts away from human-oriented browser scripting and toward agent reasoning. Rather than outputting verbose JSON or HTML that consumes tokens and requires DOM parsing, BrowserAct returns compact indexed text ("click 3 / input 2") and attaches semantic memory to each browser session for task matching. This reflects the constraint that LLMs work better with concise, structured state than with raw page markup.

The companion tool, Skill Forge, extends the utility by automating the most tedious part of integration: discovering how each target website structures its data and APIs, then codifying that knowledge into a reusable Skill. This lowers the barrier for agents to scrape at scale without hand-written scrapers.

No comments yet. Be the first to share your thoughts!