Browser automation CLI for AI agents. Use when the user needs to inspect, test, or automate browser behavior: navigating pages, filling forms, clicking buttons, taking screenshots, extracting page data, reading selected Open Design browser-tab context, testing web apps, dogfooding Open Design previews, QA, bug hunts, or reviewing app quality. Prefer local Open Design preview URLs unless the user explicitly asks for external browsing.
agent-browser is a browser-automation skill (curated from Vercel Labs' agent-browser) that lets an AI agent inspect, test, and automate browser behavior through a dedicated CLI. It solves the problem of validating web apps and Open Design previews from the agent's side: navigating pages, filling forms, clicking, taking screenshots, extracting page data (logos, fonts, colors, images, motion code, OG metadata, accessibility), and reading selected browser-tab context. It is local-first, preferring local Open Design preview URLs unless the user explicitly asks for external browsing.
The workflow emphasizes disciplined, evidence-driven use. It verifies the CLI is installed (and tells the user to install it via npm rather than substituting ad hoc scripts), practices context hygiene by redirecting large upstream guides to temp files and extracting only relevant lines, and collects the smallest useful evidence first (confirm title/URL, snapshot before extraction, prefer DOM/CSS/accessibility evidence over guessing from screenshots). A CDP startup contract governs how it attaches to an existing Chrome DevTools Protocol endpoint on 127.0.0.1:9223, checking the endpoint with curl and launching a dedicated temporary Chrome profile with remote debugging only when needed, with a cleanup trap to close that profile afterward.
Target users are developers doing QA, bug hunts, dogfooding, and design-reference extraction against local previews or web apps. The scripted operations are read/inspection-oriented against local endpoints, use a throwaway Chrome user-data directory, and avoid pasting sensitive full-page dumps into chat, keeping the footprint contained. Note that launching Chrome with remote debugging opens a local control interface, so it is best used on trusted local previews.
If command -v agent-browser fails, install it globally with npm i -g agent-browser and run agent-browser install. The skill instructs the agent to stop and ask you to install it rather than use ad hoc browser scripts.
It is local-first and prefers local Open Design preview URLs. It only browses external sites when you explicitly ask for external browsing.
It attaches to an existing Chrome DevTools Protocol endpoint (default 127.0.0.1:9223). It checks the endpoint with curl and, if unavailable, launches Chrome with --remote-debugging-port and a temporary user-data directory, then connects; it never runs agent-browser open before connect.
Yes. For the smoke path it uses a temporary Chrome profile and a shell trap that kills that profile on exit, and it uses a temp HOME and stable session name.
It practices context hygiene: full upstream guides and page HTML are saved to temp files rather than printed, and it summarizes only relevant selectors, tokens, URLs, and screenshots instead of dumping large assets into chat.
Quick Setup:
.claude/skills/Repository
nexu-io/open-design