The Agent Skills Directory

[COMMAND_EXECUTION]: The framework instructions provide documentation and examples for using ShellTools and MCPTools (via stdio transport) in SKILL.md and references/mcp.md. These tools allow the agent to execute shell commands and interact with local command-line utilities.
[DYNAMIC_EXECUTION]: The skill documents PythonTools in references/tools.md, which enables the agent to execute arbitrary Python code. This is a core feature of the framework for handling data processing and logic.
[REMOTE_CODE_EXECUTION]: Several examples in SKILL.md demonstrate using uvx and npx to run remote MCP servers (e.g., 'uvx mcp-server-git', 'npx -y @openbnb/mcp-server-airbnb'). While these are standard developer tools, they involve downloading and executing external packages.
[CREDENTIALS_UNSAFE]: The documentation contains a hardcoded database connection string ('postgresql+psycopg://ai:ai@localhost:5532/ai') used for local development examples in SKILL.md and references/learning.md.
[INDIRECT_PROMPT_INJECTION]: The skill describes building agents that ingest data from untrusted sources via tools.
Ingestion points: Tools such as MCPTools, YFinanceTools, and WebSearchTools (referenced in SKILL.md and references/tools.md).
Boundary markers: Not explicitly detailed in the reference documentation; standard framework prompts are used.
Capability inventory: Significant capabilities including ShellTools (command execution), PythonTools (code execution), and FileTools (file system access).
Sanitization: Relies on default model instructions; no specific data sanitization or escaping mechanisms are documented for external tool outputs.
[EXTERNAL_DOWNLOADS]: The skill references the official agno package on PyPI and various third-party MCP servers hosted on public registries.

agno