The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides instructions for configuring MCP servers that ingest data from external sources such as CRM records, account enrichment APIs, and customer support threads.
Ingestion points: Untrusted data enters the agent context through tools connected to CRM platforms, support systems (like Plain), and enrichment services mentioned in Phase 1 (SKILL.md).
Boundary markers: The skill explicitly mandates the use of human approval policies for write actions and recommends "read-only first" permissions as a default boundary.
Capability inventory: The configured tools have capabilities to research accounts, enrich contacts, inspect CRM data, and draft messages (SKILL.md).
Sanitization: The methodology focuses on tool curation, permission design, and audit logging to mitigate the risks of processing untrusted external data.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill mentions a remote server endpoint at https://mcp.plain.com/mcp. This is identified as a reference to a well-known customer support platform (Plain) and is documented as a standard integration point using OAuth for user-scoped permissions. The skill also includes specific warnings against hardcoding credentials, directing users to use environment variables instead.

mcp-setup