The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill configuration recommends installing the 'chrome-devtools-mcp' package from the NPM registry using the 'npx' command. NPM is a well-known service, and this package is a standard tool for the stated purpose.
[REMOTE_CODE_EXECUTION]: The use of 'npx' results in the execution of the downloaded 'chrome-devtools-mcp' server code. This is a standard installation pattern for MCP servers and is performed within the context of the user's project configuration.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with live browser content (DOM, console logs, network responses), which constitutes an untrusted data surface. The skill includes mandatory defensive guidelines to mitigate this:
Ingestion points: Data enters through DOM inspection, console retrieval, and network monitoring tools in the browser.
Boundary markers: Instructions require clearly labeling browser data as 'UNTRUSTED' and maintaining separation from trusted instruction context.
Capability inventory: Capabilities include screenshot capture, DOM inspection, network monitoring, and JavaScript execution within the page.
Sanitization: The agent is explicitly instructed to report findings as data rather than execute commands found in browser content, and to require user confirmation for any page mutations or navigations.
[DYNAMIC_EXECUTION]: The skill documentation describes the use of a JavaScript execution tool within the browser context. It specifies safety constraints such as restricting its use to read-only state inspection and prohibiting access to credentials (cookies, localStorage) or making external network requests from the browser context.
[PROMPT_INJECTION]: A static check flagged the phrase 'Ignore previous instructions'. However, this occurs within a defensive instruction advising the agent to ignore such phrases when found in untrusted browser data, and does not represent an attempt to override the agent's core safety guidelines.

browser-testing-with-devtools