The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external MCP catalogs and server candidates, creating a potential surface for indirect prompt injection.
Ingestion points: Data enters the context via the search_mcp_catalog and inspect_mcp_candidate tools as described in SKILL.md.
Boundary markers: The instructions include explicit safeguards, such as requiring the agent to "collect only the non-secret configuration fields" and strictly forbidding the request of API keys or tokens in the chat interface.
Capability inventory: The skill's tools are scoped to project management and MCP registration within the authenticated platform environment.
Sanitization: The workflow is restricted to "reviewed" candidates from the server-owned catalog, reducing the risk of processing malicious third-party content.
[SAFE]: The skill implements authentication best practices by directing users to external platform URLs for sensitive operations (OAuth/Connect) rather than attempting to process secrets within the AI's conversation context.
[SAFE]: Data exfiltration risks are mitigated through strict rules on the send_platform_mcp_feedback tool, which requires user consent and prohibits the inclusion of identifiers, credentials, or payloads.

add-mcp-from-catalog