Give your AI agent eyes and hands on the web.
Connect to your daily browser with full login state. Search, browse, click, extract — your agent operates the web like you do.
npx skills add eze-is/web-access
Works with
Built-in web tools return summaries, not source. They can't log in, can't interact, can't see what you see.
web-access connects directly to your daily Chrome via CDP. No Puppeteer, no headless browser, no separate profile. Your agent sees exactly what you see.
Query the DOM, extract text, discover links and structure
Click buttons, scroll pages, fill forms, upload files
Extract content, capture screenshots, analyze video frames
Multi-layer search: engine results → primary sources → full page extraction. Verify facts against official documentation, not second-hand reports.
> "Research the latest Claude model pricing and compare with GPT-4o"
Operates inside your logged-in Chrome. Access dashboards, admin panels, internal tools — anything you can see, your agent can see.
> "Check my Vercel deployment status and summarize recent errors"
Xiaohongshu, WeChat Official Accounts, Zhihu, Weibo — platforms that block static fetching work naturally through the real browser.
> "Grab the top 10 posts about AI agents on Xiaohongshu"
Click, scroll, type, upload files, handle popups. Navigate multi-step flows like a human — form submissions, checkout processes, configuration wizards.
> "Fill out the application form on this page with my info"
Seek to any timestamp, capture frames, analyze video content visually. Extract information from tutorials, presentations, and recordings.
> "Watch this demo video and summarize the key features shown"
Spawn multiple agents, each with their own browser tabs. Research N topics simultaneously — total time equals the slowest single task.
> "Compare these 5 competitor products — pricing, features, reviews"
The CDP Proxy exposes a clean REST API. Your agent calls curl — no SDK, no dependency, no complexity.
/new?url=...
Open a new background tab
/eval
Execute JavaScript in page context
/click
Click an element by CSS selector
/screenshot
Capture page or video frame
/scroll
Scroll to position or direction
/navigate
Navigate to a new URL
/setFiles
Upload local files to file inputs
/close
Close a tab when done
npx skills add eze-is/web-access
Open chrome://inspect/#remote-debugging in Chrome, check "Allow remote debugging", restart Chrome.
Ask your agent to browse, search, or interact with any website. The skill activates automatically.
All operations happen in new background tabs created by the agent. Your existing tabs are never touched. When the task is done, the agent closes its own tabs.
Yes. Since it uses your real Chrome with all your login sessions, it can access any site you're logged into — including corporate intranets, admin dashboards, and internal tools.
The skill operates through your real browser, not a headless automation tool. Most websites can't distinguish it from normal browsing. However, extremely aggressive anti-bot systems may still flag automated patterns.
Yes. Each sub-agent creates its own tabs and operates independently. They share one Chrome instance and one proxy, using different tab IDs — no conflicts.
web-access follows the SKILL.md standard. It works with any agent harness that supports skills — Claude Code, Cursor, Gemini CLI, Codex CLI, and others.
web-access is designed for local development where you have a running Chrome instance. For server environments, you'd need a headed Chrome session accessible via CDP.