The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs the use of npx to execute the mcp-remote package for connecting to the vendor's official endpoint at https://mcp.rudderstack.com/mcp. This is a standard and documented method for remote MCP connections provided by the vendor.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of data from external sources such as live event streams and warehouse query results, which constitutes an indirect prompt injection surface.
Ingestion points: Data is retrieved via tool calls to the RudderStack MCP server (SKILL.md).
Boundary markers: While no specific delimiters (like XML tags) are mandated for the prompt itself, the instructions tell the agent to "Extract only expected fields" from tool responses.
Capability inventory: The skill is configured with Read, Write, and Edit permissions for managing workspace resources.
Sanitization: The skill includes a dedicated "Handling External Content" section which explicitly directs the agent to sanitize warehouse results, validate identifiers, and refrain from executing any code returned by the inspection tools.
[CREDENTIALS_SAFE]: The skill uses OAuth for authentication and explicitly warns the user against committing session tokens or the mcp-remote cache directory to version control.

rudder-mcp-workflow