The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documentation proactively identifies the risk of indirect prompt injection in local MCP servers. It warns developers that tool inputs must be treated as untrusted and provides remediation strategies such as path normalization and containment checks. Ingestion points are identified as tool parameters; capability inventory includes filesystem and process access; and sanitization examples are provided using Node.js and Python path utilities.\n- [EXTERNAL_DOWNLOADS]: The skill references official tools and schemas from Anthropic, including the @anthropic-ai/mcpb utility and schemas hosted on Anthropic's GitHub repository. These downloads are standard components of the development workflow and originate from a trusted technology vendor.\n- [COMMAND_EXECUTION]: The documentation provides explicit instructions on how to avoid shell injection vulnerabilities. It specifically recommends using array-based argument passing with execFile instead of shell interpolation with exec to ensure that user input cannot lead to arbitrary code execution.\n- [CREDENTIALS_UNSAFE]: The skill describes best practices for secret management using the sensitive: true flag in the server manifest. This ensures that sensitive values like API keys are stored in the OS keychain rather than in plaintext files, and are handled securely as environment variables at runtime.

build-mcpb