The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external sources (Home Assistant device states, logbook entries, and history) which can influence the agent's behavior.
Ingestion points: Data is pulled into the agent context via tools such as states_control, history_query, logbook_query, error_log_get, and calendar_access in SKILL.md.
Boundary markers: The instructions lack explicit boundary markers or warnings to ignore instructions embedded within the Home Assistant data.
Capability inventory: The skill provides powerful tools such as call_service (generic service execution), template_render (Jinja2 evaluation), camera_control, and lock_control.
Sanitization: There is no evidence of sanitization for data interpolated into the template_render tool or passed to service calls.
[DYNAMIC_EXECUTION]: The skill provides a tool (template_render) that executes Jinja2 templates provided at runtime.
Evidence: The template_render tool in SKILL.md allows the execution of arbitrary Jinja2 logic, which could be exploited if combined with untrusted data retrieved from device states or attributes.
[EXTERNAL_DOWNLOADS]: The skill instructions direct the user to download and install external code from public sources.
Evidence: The Installation section in SKILL.md instructs users to install the homeassistant-mcp package from PyPI and provides commands to clone the source code from https://github.com/robbrad/homeassistant-mcp.git.

homeassistant-mcp-server