Until now, an AI agent connected to Home Assistant could only work with the entities you exposed to Assist. Home Assistant 2026.11, due in early November, adds a second LLM API aimed at the system itself. It is called "Home Assistant", it only works for administrator accounts, and it is not limited to exposed entities. On 18 October Paulus Schoutsen merged its two newest tools, one for history and one for templates, into the development branch.

Two APIs, two jobs

The new API arrived on 8 October in a pull request by Allen Porter, the same day the core architecture meeting approved separate LLM API scopes. The Assist API keeps controlling and reading exposed devices, and on 12 October it was renamed "Device control". Its internal ID did not change, so existing setups keep working.

The "Home Assistant" API now has four tools, all marked read-only:

  • System log. Reads warning, error and critical entries from the in-memory log, filtered by level and by logger or source path, with an optional traceback for each exception.
  • Search (merged 16 October). Finds entities, devices, areas, labels, automations, scripts and scenes by name. Results are ranked, grouped by type and capped at five per type, and the agent can page through one type.
  • History (merged 18 October). Returns long-term statistics for one or more entities as CSV, with local ISO 8601 times and the entity's unit. By default it covers the last 24 hours, and row length follows the span: 5-minute rows for up to 6 hours, hourly up to two days, daily up to 62 days, weekly up to a year, monthly beyond. Requests that would return more than 500 rows are refused. Only entities with numeric statistics (sensors with a state class) are covered. Plain state changes are not included, though the pull request says they may come later.
  • Render template (merged 18 October). Renders a Jinja2 template in strict mode, with optional variables such as sample trigger data. Rendering stops after 3 seconds and output is capped at 16 KB. The tool sits in the core homeassistant integration because the template integration only loads when you have template entities.

Together they let an agent read the log when an automation fails, find a sensor by name, summarise last week's energy use and test a template before you paste it into an automation.

Who gets access

Home Assistant refuses the API to any user who is not an administrator. Allen Porter's pull request spells out the trade-off: picking this API gives the conversation agent or MCP client access to log messages and tracebacks, which may be sent to its configured LLM provider.

That matters if you connect an AI client through Home Assistant's MCP server, which 2026.10 turned into a one-click setup. New MCP server entries offer all LLM APIs by default, so a client signed in as an administrator would get the new tools as well. The read-only switch added to the MCP server on 12 October will not hide them, since all four are marked read-only. If an external AI service should not read your logs or history, select only "Device control" in the MCP server options.