Private beta

A team of AI agents that actually does the work.

They use your computer, work in your browser, and reach your accounts. They share what they learn — and they only ever repeat what they are permitted to.

Request access See how it works

macOS today · Windows in beta · your own AI subscription

One task, start to finishlive

You chase the three overdue invoices opened QuickBooks3 overdue, $18,400 drafted follow-ups in Gmail3 drafts, unsent checked the calendarAcme call Thursday remembered for the teamAcme pays ~20d late

then, from someone else’s laptop

Dana · Financeis Acme a payment risk? → Yes — flagged Thursday. They run about 20 days late. Sam · Supportwhat is Priya’s salary band? → I don’t have anything on that.

Sam’s agent isn’t declining to answer. It was never given the record — the database refused it before the model saw it. There is nothing to talk it out of.

What it does

Agents that operate, not just answer.

Uses your computer

Opens apps, fills forms, moves files, drives a browser. Watch it work, and take the mouse back whenever you want.

Reaches your accounts

Gmail, Calendar, Slack, Drive, Notion and several hundred more in one click. Anything missing, add it yourself with an API key — QuickBooks included.

Remembers as a team

What one agent learns, the others can use — held as a graph of facts and relationships, not a folder of notes nobody re-reads.

The part that is different

The person a fact is about decides who reads it.

Most systems let whoever wrote something decide who sees it. That is backwards for anything about a person. Here, each individual sets the rule for facts about them, by category and by audience — and the company sets its own rules the same way.

One record — "Priya, compensation" — seen by four people
Who is asking	Their standing	Result
Priya	The person it is about	always
Dana, Finance	A role Priya opened	permitted
Sam, Support	No rule covers them	refused
Any agent	Asks on someone's behalf	inherits that person's answer

Enforced in the database, not in the AI's instructions. If you are not permitted to see something, the model never receives it — so it cannot repeat it, be argued into it, or leak it by accident. That is a different kind of claim from "we told the model to be careful," and it is the one that survives a bad day.

Handling sensitive data

Built for the review, not the demo.

01 Nothing runs in our cloud

The app runs on your machine and your files stay there. There is no server of ours in the middle holding your data.

02 Identifiers stripped at the boundary

Everything returned from a connected account is scanned before an agent sees it — SSNs, medical record numbers, NPI and DEA numbers, dates of birth, addresses, phone, email, account and card numbers. Redact them, or block the response entirely.

03 It fails closed

If the scanner cannot process something, it returns nothing rather than returning it unredacted. The failure mode is an inconvenience, never a disclosure.

04 Separation the database enforces

One organization's memory is unreachable from another's by the database's own rules — not by application code remembering to filter. The service refuses to start at all if given a database account privileged enough to bypass them.

On HIPAA, plainly. Compliance is not a certificate you buy — it is controls plus a signed agreement covering wherever the AI model runs. The controls described here are built and tested. The agreement depends on which model provider you use, and we would rather walk you through the data flow than quote you a date.

Getting started

Four steps, about ten minutes.

Request access

Tell us who you are and what you would put it to work on. We are onboarding a small number of teams by hand.
Install

Download and open it. Bring your own AI subscription — we recommend Anthropic, and work with any of them.
Connect your accounts

Pick the services your work already lives in. One click each; nothing is copied anywhere.
Paste your team code

One code joins your agents to your team's shared memory, with your permissions already in force.

See it on your own work.

The fastest way to judge this is thirty minutes with your actual accounts connected. We will set it up with you.

Request access