Install
get-skillsWho it's for
Look at BrowserAct when an AI agent or a team needs real-browser automation, web data extraction, or account-based flows. Official wording: BrowserAct enables AI agents and teams to perform real-browser automation, web data extraction, and account-based workflows.
The repository title also mentions a browser-automation CLI for agents, handing control to a person when stuck, parallel tasks, independent multi-session work, and isolated multi-account browsing. It fits work that already happens in a real browser, not people who only parse static HTML. The body gives no client roster; do not write support for only one agent.
Page: https://github.com/browser-act/skills
First try in 3 minutes
The body gives no npx or other installer. Official usage is: Your agent can load the BrowserAct Skill, discover browser state with get-skills, and run browser automation commands directly from your local environment. The first try stops at discovering state; do not invent an install command or a local directory.
Preconditions: the work is real-browser automation, extraction, or an account flow; and you have already decided between fully cloud-managed execution and local browser control driven by your own agent workflow.
Send get-skills to the agent and ask it only to describe the current browser state and what state would return next. The success signal is that the agent can restate state returns an indexed list, then click 3 / input 2 "..." can be used, and No DOM parsing required. If the agent clicks, types, or starts cloud execution at once, stop first.
Use cases
Case 1: Operate a page locally by index. First get-skills to see state, then call state as needed, and run click 3 or input 2 "..." with the returned numbers. Click and input change the page; confirm the numbers before running. The success signal is that the agent works by index commands, not by parsing the DOM. The page does not write more command names; do not invent them.
Case 2: Hand control to a person when stuck. remote-assist generates a live URL; a user can take over from any device, then the agent continues. This produces a reachable live address; confirm you want that URL opened before calling it. The success signal is a live URL plus an explanation that the agent can resume after the person finishes. Do not merge the two modes (fully cloud-managed / local browser control) into one step.
Usage limits
There is no copyable installer. Do not rewrite “load the BrowserAct Skill” into an install command that does not appear.
The written commands are only get-skills, state, click 3, input 2 "...", and remote-assist. Do not invent other CLI flags, anti-bot methods, or account-login steps.
You may choose only the two modes written on the page; do not invent a third.
The remote-assist URL is an outbound address; do not generate it before confirmation. Do not invent unwritten switches for parallel tasks, multi-session, or multi-account work; the title mentions those capabilities, but the body gives no matching commands.