The Agent Skills Directory

[COMMAND_EXECUTION]: The skill operates by spawning fresh subprocesses for tool invocations, specifically executing commands through npx safari-mcp.
[EXTERNAL_DOWNLOADS]: The skill is configured to automatically download and run the safari-mcp package from the npm registry upon its first use.
[REMOTE_CODE_EXECUTION]: Commands such as tool evaluate and tool run-script allow the execution of arbitrary JavaScript code within the context of the active Safari browser page.
[DATA_EXFILTRATION]: The toolset includes capabilities to access sensitive information, including browser cookies (get-cookies), local storage (local-storage), and the macOS system clipboard (clipboard-read).
[PROMPT_INJECTION]: As a browser automation harness, the skill has a significant attack surface for indirect prompt injection, where malicious instructions on a website could influence the agent's actions.
Ingestion points: Processes external data from URLs via commands like tool navigate and tool snapshot.
Boundary markers: The skill documentation references URL schema validation in utils/security.py to prevent access to dangerous protocols.
Capability inventory: Includes the ability to write files (save-pdf), execute JavaScript (evaluate), and modify the system clipboard.
Sanitization: Implements validation to block dangerous schemes such as file:, javascript:, and data:, and can optionally block private network access.

cli-anything-safari