The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads and executes the chrome-devtools-mcp package from the official npm registry using npx. This is a verified, well-known tool maintained by the Google Chrome DevTools team for browser automation tasks.
[COMMAND_EXECUTION]: The skill utilizes a local bootstrap bash script (scripts/ensure-jb-chrome-mcp.sh) to configure the environment and relies on the mcporter CLI to communicate with the browser's remote debugging interface.
[DYNAMIC_EXECUTION]: The skill exposes the evaluate_script tool, which enables the agent to execute JavaScript directly within the browser's context. This is an expected and documented feature of the underlying DevTools protocol server used for page inspection.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface for indirect prompt injection by ingesting untrusted data from active browser tabs (DOM snapshots, console logs, and network data).
Ingestion points: Browser content is accessed via tools such as take_snapshot, list_console_messages, and get_network_request as defined in the tool reference.
Boundary markers: The SKILL.md file provides explicit instructions for the agent to redacting sensitive information like passwords or auth tokens from its output.
Capability inventory: The skill can navigate pages, interact with forms, manage extensions, and execute scripts on the host via the mcporter wrapper.
Sanitization: The 'Secret handling' section instructs the agent to return only the 'shape' (presence, length, status) of tokens rather than the actual values.

jb-chrome-mcp