A MakeUseOf writer connected Claude Code to Home Assistant with a long-lived access token, letting the AI inspect the setup, call device actions and help build automations. The experiment shows how an assistant can move beyond drafting configuration to making changes that affect a real home, which is useful and rather more consequential than another dashboard mock-up.
Anthropic Watch analysis
What happened
The writer says the token gave Claude access to Home Assistant’s API, within the permissions of the account it belonged to. In his setup, Claude could read sensor states and control lights, plugs and other connected devices. He used it to build dashboards, troubleshoot a camera setup and create automations, including one that switches off devices when he leaves home.
He also describes a script for an air conditioner that checks for actual power draw, retries if the compressor has not started, and stops after five attempts. The account is a personal experiment, not a claim that every Home Assistant installation or Claude setup will behave the same way. In MakeUseOf’s account, the writer says Home Assistant permissions still applied: the token did not itself grant SSH access, and deploying code required a separate route.
Why it matters
There is a meaningful jump between asking an assistant to write YAML and giving it a credential that can operate devices. A mistaken configuration can do more than make a screen look odd; it could switch equipment on or off. The writer says he tightened permissions, made approvals more deliberate and began checking code before deployment.
His warning is specific: Home Assistant long-lived tokens can remain valid for ten years. That is a very long lease for a credential copied into an AI chat, especially when the assistant is not running on your own hardware. The article recommends a dedicated credential, limiting deployment access, backing up configurations and revoking the token when the work is done.
Our read
This is a compelling example of AI helping with the messy middle: troubleshooting integrations, translating household rules into automations and revising work when the first attempt falls short. It is also a reminder that convenience depends on the permissions you hand over. If you try a similar setup, start with a dedicated, limited credential, keep approval gates for consequential changes and verify what will be deployed. “It worked in my house” is a useful demonstration, not a substitute for testing in yours.
What to watch
- Whether Home Assistant or Anthropic provides a more direct, limited-permission connection for this kind of workflow.
- How users manage credentials and approvals when an AI agent can call device-control actions.
- Whether future accounts show these integrations becoming reliable enough for routine household use.
Discussion spark: Should an AI assistant be allowed to operate smart-home devices directly when a person can limit its permissions and approve changes, or should it only propose actions for someone to carry out?
Sources and evidence
- Source update (31 October 2026, 15:45 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.