The Agent Skills Directory

[COMMAND_EXECUTION]: The stdioTransport utility allows the agent to communicate with MCP servers by executing local commands (e.g., using npx or specific binary paths). This is a primary feature of the Model Context Protocol for interacting with locally hosted tools.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external data from MCP servers through resources, prompts, and tool results. This creates a surface for indirect prompt injection if a connected third-party server returns malicious instructions.
Ingestion points: Data enters the agent context via client.readResource(), client.tools(), and client.prompts() (SKILL.md).
Boundary markers: The skill relies on the underlying tanstack-ai framework's chat() implementation for handling tool results and message formatting.
Capability inventory: The skill possesses the capability to execute commands via stdioTransport and perform network operations via StreamableHTTPClientTransport.
Sanitization: Sanitization of external content is handled at the framework level by the tanstack-ai adapters and tool-calling mechanisms.

ai-mcp