Web Access

Problem

Many web tasks are not solved by searching and reading a static body. Logged-in dashboards, internal tools, dynamically rendered pages, collapsed content, lazy-loaded media, and video frames often sit beyond what WebSearch, WebFetch, or curl can reach. This skill targets workflows that require a real browser context for discovery, extraction, and interaction.

How It Works

It checks Node.js 22+, browser debugging availability, and a CDP Proxy, then drives a local Chromium-based browser over an HTTP API. Core capabilities include:
- Static and dynamic reading: use WebFetch, curl, or optional Jina preprocessing for known URLs; use DOM evaluation when raw structure matters.
- In-browser interaction: inspect structure and perform actions with /eval, /click, /scroll, and /screenshot, which helps with logged-in pages and anti-scraping constraints.
- Media and video: extract image URLs from the DOM, seek a <video> element, and capture frames for discrete content analysis.
- Parallel research: independent targets can be split across sub-Agents using separate targetId values, keeping large page content out of the main context.

Boundaries And Notes

This skill depends on a local browser and CDP connection, so initial setup may require user authorization. Rapidly opening many pages can trigger anti-crawl controls. It should not close existing user tabs; temporary tabs created by the skill should be cleaned up. Login prompts should be based on whether the target content is actually inaccessible. Site-pattern files should contain only verified facts, with fallback to generic behavior when they fail.