Money, documents, photos, property and health for the whole family, in one place you actually own. It runs on your machine, and nothing leaves it that you did not connect yourself — you run it by talking to your assistant: Claude or Codex.
or self-host the web app on your own server →Illustration — not an actual screenshot.
Postgres, storage and the AI models all run in Docker on hardware you control. There is no vendor account holding your family's finances, health and papers — because there is no vendor server in the path.
You don't click through forms — you talk to your assistant in Claude. 151 tools let it file receipts, answer "what did we spend on the kids in June?", open the right screen, and run the morning brief while you sleep.
Bookkeeping, documents, photos, rentals, health, school and food stop living in twelve apps and one shoebox. It's a single, searchable household — with per-person access so kids see only what they should.
Real interactions, verbatim from daily use — the assistant reads and writes the same household your family sees in the app. A fresh assistant knows nothing about your family; this one knows the receipts, the schedules and the appointments, because you kept them. Models will keep getting better — what none of them will have is your family's twenty years.
Everything below is running today against the developer's own family data. Nothing here is a mockup of a feature that doesn't exist.
Drop a receipt into the chat and get back a balanced, categorized transaction. A real ledger underneath: EÜR, USt-VA, BWA and DATEV export, multi-currency, gift-tax allowances.
Banks pull themselves in overnight: PSD2 across European banks, plus monobank and Wise. Nothing books itself — new movements wait as drafts until a person confirms them. Mail, Google Calendar, shops, school portals and messengers connect the same way, and each stays off until you turn it on.
Run a second home from the same assistant — grandparents abroad, a relative you care for. Each household stays sovereign; you share only what you choose.
OCR'd scans, passwords and 2FA in an encrypted vault. Ask in plain language; results are filtered by who's allowed to see them.
AI captions make every photo findable — "Emma's first snow" returns the photo, not a folder. Face recognition is optional and local.
Properties, leases, rental income, utility bills and meter readings — with tenant tracking and CSV import for the boring parts.
Vaccinations, medical tests and pregnancies; school, enrollments and grades; recipes, meal plans and shopping lists that write themselves.
A button backs up the database and every file to a disk you own, encrypted if you ask for it — and the same scripts ship inside the installer for when you would rather type. Restore is one command. Reports export to CSV and the books to DATEV, so nothing is trapped in a format only we can read.
A register of the accounts, subscriptions, domains and backups that matter: who inherits each, and how that person gets in. Naming someone grants them access to the entry — it is a permission, not a note in a drawer. The page leads with what you have not decided yet, and plans ask to be revisited on a schedule.
The installer does the hard bits. You don't touch a terminal unless you want to.
Download for macOS or Windows and run it. One window, a short setup wizard, no accounts to create.
The app provisions the whole stack for you — a WSL environment on Windows, a lightweight VM on Mac — and starts Postgres, storage and the AI models in Docker on your machine.
Connect Claude with one token and just ask. show me the photos from last summer, file this receipt, how much rent is overdue?
“AgentoNest isn't a startup's guess at what families need. It runs my own household every day — paperwork in two countries, school schedules, thousands of receipts. Every feature on this page exists because my family needed it, and it ships only after we've lived with it. And when my children are grown, this is what I hand over — not a drawer of dead logins, but a family record they can still open.”
— Oleksii Kulinich, founder
Not a privacy policy you have to trust — a design you can inspect.
Fine-grained permissions via OpenFGA: children can't see parents' finances or health. can_read(health) = self ∨ (parent ∧ minor) ∨ admin.
Passwords, 2FA seeds and sensitive fields sealed with AES-256-GCM. Secure-by-default: without authorization, the answer is no.
Every change is logged. One command backs up the database and every file to a disk you hold, encrypted if you want; one command puts it back. Both scripts ship inside the installer — nothing to buy, nobody to ask.
Manage grandparents' household remotely without merging data — each home stays sovereign, shared only where you choose.
You don't file a ticket in a portal. You say it out loud to your assistant — log this in the fix queue — and it writes the card for you, with the version, the screen and what you were doing already filled in.
A broken screen, a wrong number, a module you wish existed. Plain words are enough — the assistant turns them into a card and files it on your own machine.
Before anything is sent you are shown the exact text, in full. Cards are written locally first, so there is always something to read and edit — and a reminder to keep passwords, account numbers and documents out of it.
Nothing is transmitted until you confirm. Say no and the card stays on your disk as your own note. Say yes and it reaches the maintainer, who reads it, keeps it while it is open and up to twelve months after, and deletes it sooner if you ask.
This is a service, not a promise. A request may be built, built differently, or not at all, and there is no obligation to reply. What sending grants us is in the Beta Terms; what actually travels is in the Privacy Notice.
No. The desktop app provisions everything itself — WSL2 on Windows, a lightweight VM on Mac — and keeps it updated. You never open a terminal unless you want to. The self-host path exists for people who enjoy terminals.
A reasonably modern laptop or desktop with 8 GB+ RAM for the local setup, or a small VPS (~€7/mo) for the always-on web version the whole family reaches from their phones.
The beta is free for invited households — every domain, every tool, nothing held back. Pricing after the beta is not settled yet, and we will not pretend otherwise. What is settled: your data lives on your hardware, and a full backup you can take at any moment means walking away costs you nothing. See the Beta Terms.
You do not connect it by hand. The installer asks which assistant you use and wires it up itself — it writes the config file where that program expects it, or hands you one command to paste where it cannot. Nothing here needs copying from a web page.
| Assistant | How it connects |
|---|---|
| Claude Desktop | set up for you — config written automatically |
| Cursor | set up for you — config written automatically |
| Antigravity (Google, Gemini) | set up for you — config written automatically; the IDE and the agy CLI share one entry |
| Claude Code | one command, shown ready to paste |
| Codex (OpenAI) | one command, shown ready to paste — the terminal CLI and the Codex extension in Cursor / VS Code share it |
| ChatGPT Desktop | set up for you — the app reads the same Codex config, so one connection covers both |
| Anything else that speaks MCP | the same server definition, yours to copy |
Which model that gives you. Claude Desktop and Claude Code run Anthropic's Claude. Codex and the ChatGPT app run OpenAI's GPT. Antigravity runs Google's Gemini — that is the route to use the family server from Gemini. Cursor lets you pick. So a household keeps whichever subscription it already pays for.
It works because the server speaks MCP, an open protocol — not because of a deal with any vendor. Every assistant above runs on your own machine and talks to your own server, so nothing has to be published to the internet for them to reach it.
Talking to an assistant spends that assistant's own usage allowance, and a household session is not small. The app also works entirely without one: every job can be done in the browser.
All three, and you keep the subscription you already pay for. Claude comes in through Claude Desktop or Claude Code, GPT through the ChatGPT app or Codex, and Gemini through Google's Antigravity — the installer sets each of them up for you. Cursor lets you choose the model yourself.
Only when you ask your assistant something, and only what answering it needed — not your database. AgentoNest has no cloud of its own, so we never receive anything and nothing syncs anywhere by itself. The assistant runs under your own account, on your own terms with that vendor. Prefer that nothing leaves at all? Use the browser interface — it does every job without an assistant.
It's your server and your data in ordinary formats — nothing stops working. Your records sit in Postgres and ordinary files, and the backup and restore scripts come with the app, so a copy of everything is always one command away. Reports export to CSV, and the books to DATEV.
The desktop app bundles the entire stack — pick your platform. Prefer to run it on a server you rent? Self-host the web app.
Native app that provisions a local VM and runs the full stack on your Mac.
One-click installer; runs the stack inside WSL2. Unsigned for now — Windows SmartScreen may warn on first run.
Run it on a small server (sealed off the internet over Tailscale) and open it in any browser at your own domain.
All links always point to the latest release on GitHub.