Reading
T0Looks without changing anything. Runs at once.
For the household’s IT adviser: the parts, the rules they follow, and where your data lives.
Every screen talks to the same steward. Every action passes through the same executor. Every connection reports into the same stream of events.
A spoken question, a tap on a tile, a routine at dusk and an errand on its third day are all the same thing underneath: a call to a tool.
By voice, by a tap, by a routine or by an errand. Each becomes a tool call.
Against your role, the screen’s limits, the space it belongs to and the day’s budget.
Anything consequential waits five minutes for your yes. The highest tier also needs your passkey.
The connector does the work, with a key opened from the vault for that one use.
Approved, declined or refused, every request is written to an append-only record.
Every tool carries a tier. An owner may raise one; nobody can lower it below its default.
Looks without changing anything. Runs at once.
Lights, scenes, drafts. Done at once, then announced.
Sending, posting, unlocking, deleting. Waits for an adult’s approval, stated in one plain line.
The alarm, and money when it moves. Approval plus the owner’s passkey from the last two minutes.
Any tool whose name says send, pay, unlock, delete or the like must be at least T2. Two automatic checks fail the build otherwise.
Children are capped at T1, guests at T0, Alexa at T1, and no AI app ever holds T3. Above its cap, a request becomes an approval, not an action.
By default no household runs more than 500 actions a day beyond reading. Routines and errands are approved once, at the tier of their riskiest step.
No endpoint and no tool returns a secret. The secrets API answers with descriptions only, and secrets are scrubbed from anything the browser’s AI sees.
The trust boundary, stated plainly: the database alone holds only ciphertext. With the service’s master key, a secret can be opened, which is how a connection keeps working while you sleep. Each opening is listed in a hash-chained record that can be verified end to end.
Work is routed by slot, so any provider can fill any slot. The model we provide is the default and the fallback; your own key takes precedence wherever you set one.
The wake word and the voice detection run on the screen itself. Speech is streamed, answered and spoken back, and you can interrupt it at any point.
AI apps connect through a full OAuth 2.1 server: metadata discovery, dynamic registration, PKCE S256 only, device flow, revocation, rotating refresh tokens. A grant acts as one member, under that member’s role.
Mount any MCP server. Tools it marks read-only run at T0, unmarked tools ask first, and destructive tools sit at T3. The owner may mark a tool read-only.
Your own tools receive an HMAC signature with every call. Inbound webhooks and Alexa requests are verified before anything is read.
A household’s region is set when it is created, in the UK or the EU. It is never taken from a request, and it does not change.
Every service reports in every 30 seconds. A public status page says what is up, what is down and what is being done about it.
99.9%
of requests answered without a server error, with 95 in 100 inside half a second.
1.5 seconds
from the wake word to the first word of the answer, for 95 in 100 questions.
Point in time
Database history that can be restored to a point in time, with a quarterly restore drill, and versioned file storage.
Tell us a little about your house and your business. We will be in touch.
A few lines about your house and your business. We read every request and reply by email.