The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches the agent-browser package and Chromium browser binaries. These resources originate from Vercel Labs, which is an established and trusted organization in the developer ecosystem.
[COMMAND_EXECUTION]: The skill automates browser tasks by executing the agent-browser CLI through shell commands and Node.js child processes. This includes the management of browser sessions and page interactions via a command-line interface.
[DYNAMIC_EXECUTION]: Includes the ability to execute JavaScript code within the browser context via the eval command. This is used for advanced testing scenarios such as verifying application state or interacting with complex UI components, but allows for arbitrary code execution within the browser instance.
[INDIRECT_PROMPT_INJECTION]: The skill operates by ingesting and acting upon content from external websites, which creates an attack surface for indirect prompt injection where malicious site content could attempt to influence agent behavior.
Ingestion points: Accessibility tree snapshots and page content from target URLs extracted from the browser session via agent-browser snapshot and agent-browser open commands.
Boundary markers: The skill uses a structured ref system (e.g., @e1, @e2) to isolate elements within the accessibility tree, providing structural separation from raw HTML content.
Capability inventory: File system writing for screenshots and PDFs, network interaction via the browser, and JavaScript execution in the browser context.
Sanitization: Interaction is performed through specialized CLI commands that target specific elements, rather than raw string interpolation of web content into shell commands.

e2e-agent-browser