typed AI workspace · early access
Talk to AI.
Keep a database.
An assistant where every answer is typed data — schemas you can see, edit, extend, and query across every conversation you've ever had.
Start with one of these
- “Find 10 well-reviewed ramen places in Tokyo”10 typed Place rows — name, ward, rating — not a paragraph
- “Extract the action items from these meeting notes”one ActionItem object per task, with owner and due date
- “Track the books I'm reading with a rating column”a Book schema you can edit, then add to
Each one lands in the composer, ready to send.
what happens when you press send
Every answer is a row.
Extractions land in a real store as typed objects, not as text you'll scroll past. Browse them as a table, cards, a map or a chart.
Place objects in a store you can browse, sort, map and query later.then say
“Add a price column.”
Chat output evaporates. Here it compiles. The schema gains a field, every row you already collected is backfilled, and the answer you got last week is still queryable this week.
ask in words
Get a filter, not a guess.
Questions compile to a query you can see and edit — the predicate appears before the rows move, so you always know what was actually searched for.
know who wrote what
Your correction wins.
Every cell records whether you or the model put it there. Fix one and it's yours — permanently, against every later pass.
And it listens.
Turn the mic on and keep talking. It works ahead of you on what you said, shows you every assumption it made as a chip you can revoke, and asks a sharper question each time your answer set narrows. Talk over it and it stops.
Start with a sentence.
Sign in with Google and your first prompt is already waiting in the composer.
Try “Find 10 well-reviewed ramen places in Tokyo”