Source: github.com/homeassistant-ai/ha-mcp (MIT, about 7.9k stars as of 2026-11-18); highlighted by André Lindenberg on LinkedIn (2026-11-17).
What it does
ha-mcp is an MCP server that gives Claude, ChatGPT and other MCP clients about 86 tools across Home Assistant: fuzzy entity search and config search, service calls and bulk device control, real-time states, automations, scripts, helpers, dashboards, areas, zones, calendars and blueprints, history and statistics, camera snapshots, automation traces, and backup, restore and updates. The flagship example is debugging: tell the assistant a motion-sensor automation isn't working and it reads the automation's execution trace, finds the failing step and edits it.
Deployment options
Recommended: an in-process integration installed through HACS, or the Marketplace on Home Assistant 2026.11+, running the full server inside Home Assistant on any install type.
Alternatives: a Home Assistant app (add-on) with no token to manage, Docker or PyPI over streamable HTTP, and local stdio (discouraged; it has known transport issues).
Remote access goes through a Home Assistant webhook URL (via Nabu Casa or an existing reverse proxy), or stays local-only with no webhook registered.
Authentication: by default the secret URL itself is the credential. Optional modes require a Home Assistant sign-in or an external OIDC provider; with OIDC, all authenticated users share the server's Home Assistant credentials.
The server acts as a chosen administrator account, either through a long-lived access token or a dedicated account it creates.
Safety model
Automatic edit backups: a snapshot before each edit, which the assistant can restore.
Read-only mode toggle and per-tool enable or disable.
Tool security policies that require user approval, but these start switched off.
File and YAML editing tools are opt-in feature flags, off by default.
The defaults lean toward capability: an admin-level credential, approvals off, and a secret URL as the default credential for remote access. That suits a hobbyist's own home. A shared or critical deployment needs to tighten them.
Why it is interesting beyond smart homes
It is a mature, heavily used example of an agent with write access to a physical system, and it shows which controls a real project converged on: snapshots before edits, read-only and per-tool switches, opt-in for the most dangerous file-level tools, and traces as the debugging substrate. Reading execution traces before editing is the same diagnose-then-change discipline that works for agents on any production system.
Risks and failure modes
An admin token in the server means any compromised client or leaked URL has full control of the home, including locks and alarms where those are integrated.
Approvals off by default means a bad plan executes without a human check.
A snapshot restores configuration, not physical side effects (a door unlocked, a heater left on).
Running more than one install method for the same client causes conflicts; the README warns to pick exactly one.
Prompt injection through entity names, calendar entries or other data the assistant reads could steer actions.
Review checklist for an agent with write access to physical systems
Is the credential scoped to what the agent needs, rather than administrator?
Are approval policies on for actions with physical or security effects?
Is there a read-only mode for exploration and debugging?
Do backups cover configuration, and is there a plan for actions that can't be undone?
Is remote access behind real authentication, not just a secret URL?
Are the agent's inputs (device names, notes, calendars) treated as untrusted?
Related
R117: an agent with write access to a smart home: plan doc, house rules, results.
R091: FastMCP serving agent skills from the MCP server.
R030: kernel-enforced sandboxes and verified policy for agents.