The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill recommends executing the @mcp-b/chrome-devtools-mcp package using npx, which involves downloading and running code from the NPM registry at runtime.
[DYNAMIC_EXECUTION]: The skill implements an evaluate_script tool that allows an agent to execute arbitrary JavaScript within the browser. Furthermore, the provided add_todo_item example uses innerHTML to inject potentially untrusted data directly into the DOM, which can lead to script execution if the data contains malicious payloads.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and act upon data from external, untrusted websites, creating a vulnerability to indirect instructions.
Ingestion points: The agent reads the browser DOM, document titles, and external API responses (e.g., the OpenWeatherMap example) via tools like list_webmcp_tools and evaluate_script.
Boundary markers: None identified in the instructions or code examples to distinguish between safe data and instructions.
Capability inventory: The skill includes powerful tools for browser interaction including evaluate_script, navigate_page, click, and fill (defined in SKILL.md).
Sanitization: The skill lacks sanitization or escaping mechanisms for the data it processes before passing it to other tools or reflecting it in the UI.
[EXTERNAL_DOWNLOADS]: The installation process involves cloning a repository from GitHub (WebMCP-org/chrome-devtools-quickstart.git) and installing dependencies via npm install.
[COMMAND_EXECUTION]: Instructions include multiple shell commands for configuring AI clients and adding MCP servers (claude mcp add, npx mcp-remote).

webmcp-chrome-devtools-quickstart