eze-is/web-access

What it solves

Many AI agents lack sophisticated web browsing capabilities, often relying on basic search or fetching tools that cannot handle dynamic content, user authentication, or complex interactions. This project provides a comprehensive "Skill" that adds advanced web access, browser automation, and strategic browsing logic to agents, allowing them to interact with the web as a human would.

How it works

The project implements a CDP (Chrome DevTools Protocol) Proxy that connects directly to a user's existing browser (Chrome or Edge) via WebSocket. This allows the agent to leverage the user's active login sessions and cookies. It provides a set of HTTP APIs for the agent to perform actions like creating new tabs, executing JavaScript, simulating real mouse clicks, uploading files, and taking screenshots. Additionally, it includes a find-url.mjs script to search through local browser bookmarks and history to find internal or previously visited pages.

Who it’s for

Developers and users of AI agents that support the SKILL.md standard (such as Claude Code, Cursor, Gemini CLI, and Codex CLI) who want their agents to have full, authenticated web browsing and automation capabilities.

Highlights

• CDP Proxy Integration: Connects to Chrome or Edge to use existing login states and handle dynamic pages.
• Automated Tool Selection: Automatically chooses between WebSearch, WebFetch, curl, Jina, and CDP based on the scenario.
• Local Resource Retrieval: Searches local browser history and bookmarks to find URLs that are not publicly searchable.
• Parallel Execution: Can dispatch sub-agents to research multiple targets in parallel while sharing a single proxy.
• Site Experience Accumulation: Stores and reuses operational experience (URL patterns, platform traits) per domain across sessions.
• Advanced Interaction: Supports JS clicks, real mouse events, and file uploads.