We don't really know most of the people around us, and it is harder than it should be to change that. The people you meet once, at a party or a course or somebody's kitchen table, get on with immediately, and then never see again. You swap numbers, you say “we should do something sometime,” and that is where it ends. There is nothing in particular to send them, no plausible reason to get in touch, and arranging one evening costs more effort than either of you will spend on a person you barely know.
The same thing happens, more slowly, to people we already know and have stopped keeping up with. Neither failure is emotional. Both are administrative.
01 Two things that broke
The first is what software does with attention. Most of the products that now hold our social lives make money from the time we spend looking at them. If the metric is time on screen, the content that wins is the content that holds you: the performed version of someone's life rather than the ordinary one, the argument rather than the agreement, the comparison rather than the update. Everyone involved is behaving reasonably. The result is toxic, and it makes other people's lives look like a competition you are losing.
The second problem is quieter. Keeping a friendship alive now costs real logistical work. You have to remember that someone mentioned a trip, follow up before it's too late, find a slot four calendars agree on, and carry a group chat that produces forty messages and no date. That work is invisible and unevenly distributed.
Care rarely fails because people stop caring. It fails on admin.
02 The people in between
Messaging apps serve the five or six people you speak to every week. Public platforms serve an audience of several hundred, most of whom are not really in your life. Between them sits a much larger group: people you like, don't see often, and would happily see again. Very little is built for them. Two things go missing there.
The first is small information. You are thinking about learning to sail, your knee is finally better, you're in Lyon for a week. None of that justifies a message to someone you speak to twice a year, and none of it is something most people want to announce publicly. So it doesn't travel, and the two of you never find out that you both sail, or that one of you lives in Lyon.
The second is plans. An idea becomes an event only when somebody puts a date on it and asks people directly. Most ideas die between those two steps.
03 What we built
ANToinette runs on that gap. You say one thing, usually as a voice note, whenever you feel like it. It comes back split into separate items: a plan with a date and a location, or an update. If needed, you can edit the wording, change who each item goes to, attach a photo, or discard. Nothing leaves until you approve it.
At 7 p.m., each of your friends receives one brief. It holds the invitations they've received, with accept, maybe and decline in the card; what friends have shared; and sometimes a suggestion to check in on someone. Then it stops. You can scroll back through previous days, but nothing new is being produced to keep you there, and the app does not interrupt again until the next evening.
You can also skip the voice note: add ANToinette's guest address to an event in your own calendar and it arrives in the app as a private draft, ready to share.
04 Who sees what
Every item carries an audience. There are three tiers, and each item is stamped with the narrowest tier that should see it, while each friendship carries the tier you've granted that person.
Privacy is enforced in the architecture rather than in the settings screen. Content moves only when both sides permit it: the sender's grant and the receiver's own filter. When the parser is unsure which tier an item belongs to, it picks the more private one. You can read the profile written about you, export everything, and delete the account. When you look for friends, contacts are matched in transit and the contact list is never stored or processed outside of our servers.
The inference runs on Vertex AI in European regions and is not used to train anyone's models.
05 The colony
This is where the ant queen comes from. You only ever speak to ANToinette: one voice note, one recipient, and no thinking about who any of it is for. Everything after that is her colony. The note is split into separate posts, then one ant works out what each one is and files it under the right category, one carries an invitation out to the guests, one carries an update to the people allowed to see it, one keeps an eye on a friend who has started to sound low and tells you.
All the work for social admin is done by the colony.
06 Where this fits
Social infrastructure — how people find each other, stay in contact, and arrange to be in the same room — has been built almost entirely by companies whose revenue depends on people not leaving. For something this basic, that is a strange arrangement. There is room for software whose success condition is time spent elsewhere. Europe, where the working assumption is that data belongs to the person it describes, is a reasonable place to try building it.
ANToinette is in both app stores. It is used every day by a growing number of people, most of whom tell us when it gets things wrong, and it is funded by the two of us. It is a long way from finished.
The question we're actually testing is whether software can help maintain friendships without becoming the place where the friendship happens. We think it can.
