The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process user-supplied remote URLs to register external MCP servers. This creates a surface where a malicious remote server could provide tool definitions or responses containing instructions designed to influence or override the agent's behavior.
Ingestion points: User-supplied remote MCP server URLs are ingested in SKILL.md and processed via tools like inspect_mcp_candidate and register_remote_mcp.
Boundary markers: The workflow requires calling inspect_mcp_candidate and presenting the "bounded evidence" (URL, tool count, authentication posture) to the user for explicit confirmation before any registration occurs.
Capability inventory: The agent has the capability to call various tools from the newly registered server using platform tools such as call_platform_tool, execute_tool, and describe_tools.
Sanitization: The skill relies on platform-side inspection (inspect_mcp_candidate) and human-in-the-loop review of metadata rather than automated sanitization of the remote server's content.
[REMOTE_CODE_EXECUTION]: The primary purpose of the skill is to register remote service endpoints and execute the tools they provide. This allows for the execution of logic defined at an external, user-provided location.
[COMMAND_EXECUTION]: The workflow enables the agent to execute tools on the registered remote servers through platform tools such as call_platform_tool and execute_tool.

add-mcp-from-remote-url