The "Home Assistant" LLM API, the admin-only toolset for AI agents that arrives in Home Assistant 2026.11, is no longer read-only. On 27 October at 20:45 Polish time, Franck Nijhof merged a pull request by Paulus Schoutsen that adds a tool for calling any action. A day earlier a second tool landed that lets the agent search for actions first.
Two new tools
Search actions (merged 26 October) matches a query against action IDs, names and descriptions and returns the three best matches with their targets and fields. If more actions match, it asks the agent for a more specific query. Field selectors that are too large are left out of the answer.
Call action (merged 27 October) runs any Home Assistant action and returns the response if the action has one. The target can use entity, device, area, floor and label IDs, and values in target and data can be templates, rendered the same way as in a script. An undefined variable, a template that takes too long, or an action that doesn't exist returns an error to the model so it can correct the call instead of failing silently.
Together with the render template tool, the API now has homeassistant__render_template, homeassistant__search_actions and homeassistant__call_action in the core homeassistant integration, next to the log, search and history tools covered on 18 October.
What this changes for access
The earlier tools could only read. Call action can do anything an administrator can do with an action: turn on lights, unlock a door, run scripts or restart Home Assistant. It is not limited to entities exposed to Assist. That is the point of the API, which is meant for managing the system itself, and it is still available only to administrator accounts.
The tool declares no safety annotations, so Home Assistant treats it as one that writes, can be destructive and reaches outside the system. That has a practical effect on the MCP server. Its read-only option, added on 12 October, only offers tools marked read-only, so it hides call action. Our 18 October article noted that the switch did not hide anything in this API, because all its tools were read-only. With call action added, it now does.
If you connect an external AI client to Home Assistant through MCP with an administrator account and have all LLM APIs enabled, check the MCP server options after updating to 2026.11. Turn on the read-only option if the client should only look, or select just "Device control" if it should only reach exposed devices.