Your budget, in the assistant you already use

If you already talk to an AI assistant, there is an obvious thing to want from it: to ask what you have left and get an answer from the budget you actually keep. Not a report, not a spreadsheet — a question, in the middle of a conversation about something else.

Now you can. Rudopo runs a server for the Model Context Protocol, the standard assistants use to connect to other apps. Add Rudopo as a custom connector in an assistant that supports remote servers, sign in when it asks, and it can read your spaces, your accounts, this month's envelopes and your transactions, and tell you what needs attention. The connection starts there, at reading, and stays there until you say otherwise.

Nothing changes without you

The part we spent our time on is what happens when you want the assistant to do more than read. You might ask it to tidy last week's uncategorized transactions, or to post the salary that just landed. Whether it may, and how, is a choice you make on a Rudopo screen, never in the chat, and an assistant cannot raise its own level.

The level most people want is Propose changes. Every change the assistant suggests lands in a review inbox in Rudopo as a proposal: the exact change, shown as the ledger would show it, with each field before and after for an edit. You accept it, fix a detail and then accept it, or dismiss it with a one-line reason the assistant can read back. Until you accept, nothing in your budget has moved. An open proposal expires on its own after a month.

If you trust a connection, you can let it make changes directly. Even then it works within limits you set: how large a single change may be, how much in a day, how many rows at once. The defaults are drawn from your own spending, so a limit is never unlimited. Anything over a limit, anything that looks like the same change twice, and any delete you have not explicitly allowed goes to the inbox instead, with a badge that says why. Every change is logged under the connection's name, and an edit can be undone for about fifteen minutes, the same as your own.

What it will never do

Some things are never offered to an assistant at any level. It cannot touch your account, your sign-in, your plan or your sharing. It cannot create or delete a space, change a space's currency, close an account or reconcile one. It cannot change a reconciled transaction's amount, date or account. It never gets a full export; it reads one page at a time. And it cannot move money, because Rudopo does not move money.

Your conversation never reaches Rudopo either. The assistant sends one request at a time — a question, or a change it proposes — and Rudopo answers it and keeps a record of what it asked to change. What you said stays with the assistant, under its own policies.

For the people who write scripts

The same door is open to your own code. The API now writes as well as reads: transactions, categories, payees, targets and the budget month, with a token at the same three levels. Every write takes an idempotency key, so a retry can never land twice, and every write can be sent as a dry run first, which answers exactly what would happen and changes nothing. The developer page has working examples in curl, Python and TypeScript, and every example is a dry run until you say otherwise.

What is next

Rules, recurring schedules and reminders are not reachable from an assistant yet; they are next. After them, a sandbox space to test against, and a way for your own tools to be told when something changes. No dates, as usual — we would rather tell you when each is real.

How to connect, what an assistant can do, and the questions people ask before they allow one, are on the assistant page. If something does not behave as this post says, tell us in #bug-reports on Discord, or through Help & Feedback in the app.

— The Rudopo Team