The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements a child process transport via MCPStdioClient, which launches and communicates with programs using a command array. This provides the agent with the capability to execute local executables.
[EXTERNAL_DOWNLOADS]: The documentation includes examples that use npx to download and run packages from the NPM registry (e.g., @modelcontextprotocol/server-github). This represents an external dependency download from a well-known service.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a significant ingestion point for untrusted data by fetching tool results and resources from external MCP servers.
Ingestion points: Data is fetched via send() (tool calls), readResource() (resource access), and listTools() (discovery).
Boundary markers: The implementation does not enforce specific delimiters or "ignore instructions" wrappers on content received from servers.
Capability inventory: The skill provides both network access (MCPHttpClient) and process execution (MCPStdioClient).
Sanitization: No automatic filtering or sanitization of external server responses is performed before returning data to the agent.
[DYNAMIC_EXECUTION]: The MCPStdioClient allows for runtime configuration of child processes, including environment variables (env) and working directories (cwd). While the documentation explicitly warns against building commands from user input, this capability provides a mechanism for dynamic process execution.

bx-ai-mcp