The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions for obtaining the chrome-devtools-mcp package. This is an official utility from the Chrome DevTools project designed to connect AI agents to browser instances via the Model Context Protocol.
[REMOTE_CODE_EXECUTION]: Recommended installation procedures for the toolkit involve using npx to download and execute the chrome-devtools-mcp package, which is the official method for utilizing this utility.
[COMMAND_EXECUTION]: The skill utilizes shell heredocs to execute Node.js scripts through the ego-browser utility. This allows the agent to perform multi-step browser interactions and complex state inspections in a single operation.
[DATA_EXFILTRATION]: Instructions are provided to read the ~/.config/opencode/opencode.json configuration file. This is intended for environment verification to ensure the browser debugging tools are correctly configured for the host platform.
[PROMPT_INJECTION]: The skill processes untrusted external data from web pages, which represents a surface for indirect prompt injection.
Ingestion points: Browser accessibility trees (via snapshotText()), results of runtime JavaScript execution (via js()), and network/console event streams (via drainEvents()).
Boundary markers: None observed in the provided interaction recipes; the skill does not explicitly define delimiters for untrusted web content.
Capability inventory: The toolkit provides high-level control over the browser, including script evaluation, raw protocol access (cdp()), and full interaction capabilities (click, type, navigate).
Sanitization: No specific content filtering or sanitization logic is described for the data retrieved from investigated web pages.

browser-debug-toolkit