BrowserAct Skills
by browser-act
BrowserAct Skills gives an agent a CLI for real-browser automation that gets through anti-bot walls, runs isolated multi-session tasks in parallel, and hands off to a human when stuck.
Install the skill
Tell your agent: "Install browser-act. Skill source: https://github.com/browser-act/skills/tree/main/browser-act"
About
This is a command-line automation layer built specifically for AI agents rather than human testers: an agent opens a browser session, calls state to see the current page's clickable elements by index, and drives it forward with click <index> and input <index> <text> instead of hand-writing CSS selectors. Two browser modes are available: a stealth engine for sites with anti-bot defenses and a real Chrome mode for using your own logged-in sessions. stealth-extract pulls structured content off pages that would otherwise block a scraper.
Sessions are isolated by a --session flag, so an agent can run several independent tasks in parallel, each with its own cookies, proxy, and fingerprint, without them bleeding into each other. When a task hits something the agent genuinely can't solve, such as a hard CAPTCHA or a login flow requiring a real human, remote-assist hands the live browser over to a person for a manual takeover, then returns control to the agent once it's past the blocker.
Key features
- CLI-first browser automation designed for AI agents, not human test authors
- Stealth browser mode for anti-bot defenses, real Chrome mode for logged-in sessions
- Index-based page interaction (state/click/input) instead of hand-written CSS selectors
- Isolated multi-session support with independent cookies, proxies, and fingerprints
- Structured data extraction from protected pages via stealth-extract
- Human handoff (remote-assist) for CAPTCHAs and blockers the agent can't clear alone
Use cases
- Scraping structured data from sites with aggressive anti-bot protection
- Running several independent, logged-in browser sessions in parallel for multi-account workflows
- Automating a multi-step web task, like a form fill or checkout, via index-based clicks instead of selectors
- Falling back to a human for a hard CAPTCHA mid-task without losing the agent's session state
Available tools
stealth-extract
Extracts structured content from a page protected by anti-bot defenses.
browser open
Opens a new browser session, stealth or real Chrome, at a given URL under a session ID.
state
Returns the current page's clickable/interactive elements, each addressed by index.
click / input
Clicks an element or types text into a field, addressed by the index returned from state.
solve-captcha
Attempts to resolve a CAPTCHA blocking the current session.
remote-assist
Hands the live browser session to a human for manual takeover when the agent is stuck.
Frequently asked questions
How does an agent interact with page elements without writing CSS selectors?
The state command returns the visible clickable elements with an index number; the agent then calls click <index> or input <index> <text> against that index instead of a selector.
What happens if the agent hits a CAPTCHA it can't solve?
It can try solve-captcha first, and if that fails, remote-assist hands the live session to a human for a manual takeover, then returns control to the agent afterward.