The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions to install the homeassistant-mcp package via PyPI and uvx, as well as cloning a repository from github.com/robbrad/homeassistant-mcp.git.
[COMMAND_EXECUTION]: The template_render tool allows the agent to execute Jinja2 templates. While this is a standard Home Assistant feature, it provides a powerful execution environment that could be exploited if the agent is influenced by malicious data.
[DATA_EXFILTRATION]: The camera_control tool supports a snapshot action that writes files to the local filesystem (e.g., /config/www/snapshots/). This file-writing capability could be misused to write data to unintended locations if the paths are not strictly validated.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes data from external devices and services.
Ingestion points: Data is ingested from Home Assistant entity states, attributes, logbook entries, history, and calendar events via tools such as states_control, history_query, and calendar_access (defined in SKILL.md).
Boundary markers: There are no explicit boundary markers or instructions to the agent to disregard instructions embedded in the device states or log data.
Capability inventory: The skill provides the agent with the ability to call arbitrary services (call_service), render templates (template_render), and write files to disk (camera_control snapshots) as seen in SKILL.md.
Sanitization: No sanitization or validation of data retrieved from Home Assistant is mentioned in the skill instructions.

homeassistant-mcp-server