Summary

Dashboard changes you can preview before saving and undo after, instead of hand-edited YAML.

Every card and view tool accepts a dry run, and a backup is written before each write and listed by id, so reverting is a call rather than a git history you never kept. The read side splits along retention: raw history answers what happened last night, while recorder statistics answer what this cost in January.

What it is

An MCP front end for a single Home Assistant instance, covering entity state and control, the recorder's history and statistics, the error log, and live editing of storage-mode Lovelace dashboards over Home Assistant's WebSocket API.

What you get

Entity state on demand, trimmed to the fields you name or returned in full, and on, off or toggle on any entity
Four ways to find an entity without knowing its id: by domain, by area name matched case-insensitively, by free-text search, or from a whole-instance overview with per-domain counts and state distributions
The low-level route out: call any service in any domain with your own data payload, list every automation, and restart the instance
Live Lovelace editing — add, replace, remove and reorder cards, add, update and remove views, and target a section inside a sections-type view — with a dry run available on every write and the change appearing instantly in open browsers
A backup written automatically before each dashboard write, listed by id and restorable, where the restore is itself backed up first
Two depths of history: raw state changes over the last N hours or an explicit window, and recorder statistics as mean, min and max buckets that outlive the short-term retention window
The error log filtered by level, integration, search term or line count, with error and warning counts computed over the filtered output
Guided prompts for the recurring jobs, such as walking through creating an automation

Requirements

A Home Assistant instance with a long-lived access token, given as HA_URL and HA_TOKEN, and either Docker — the published image is the recommended path — or Python 3.13 with uv. Getting HA_URL right is the usual first stumble: from a container on the same machine that is host.docker.internal rather than localhost. Saving a dashboard needs an admin token and works only on storage-mode dashboards, so a YAML-mode one stays readable but not writable. Under Docker, the backup directory has to be volume-mounted for backups to survive the container.

Setup effort

One command plus a key — uvx hass-mcp, then supply credentials