Skip to content
SageFin

Connecting an AI app to your data

SageFin can let another app you already use — Claude, Cursor, or something you wrote yourself — read your financial data and answer questions about it. You would do this to ask things in a tool you already have open, or to script something SageFin does not do itself.

It is off until you create a token or sign an app in, and it is worth understanding what you are agreeing to before you do.

This is not the assistant inside SageFin

SageFin has its own chat, and it is a different arrangement. That assistant runs on our side, under the terms in our privacy policy, and what SageFin does with what it finds describes exactly what it sends and when.

What this page is about is the opposite direction: you connecting an app we do not run.

What leaves your household

This is the part worth reading twice, because the answer is genuinely different from everywhere else in SageFin.

When you connect an outside app, the data it reads goes to that app's model, under that company's terms. Your transactions become part of a conversation living in their history, on their retention schedule. Our privacy policy does not reach it, and neither do we — once a request is authorized and answered, we cannot see what the other end did with the answer.

That is not a warning against doing it. It is the same trade as connecting anything to anything, and the tools are worth having. But it deserves a deliberate decision rather than a ticked box, and the question to ask is the one you would ask of any app: do I trust this one with my bank history?

Two things stay true regardless. Nothing is shared with other SageFin households — every request is scoped to yours. And a token can never reach the admin surface, because that permission comes from your sign-in and a token does not carry it.

Read-only until you say otherwise

By default an outside app can read and nothing else.

The write tools — setting a transaction's category, tagging one, hiding one, assigning one to a business, creating a manual expense, handing over a retailer order page, creating, renaming or deleting a tag, creating or renaming a category, setting a category's budget for a month, adding a manual account and updating it from its statement or with its value, and choosing a manual account's institution — are off behind two independent gates, and both must be open:

Either one closed means every connected app is read-only, whatever the other says.

With both open, a write happens when the app asks for it. SageFin does not stop and ask you first, because the only one it could ask is the app. The check that involves you is the app's own: most ask before they use a tool that changes something, and SageFin labels each of its tools as one that reads or one that changes, so the app knows which to ask about. If yours offers to stop asking, that is the setting that decides how much it can do unattended.

A change an app makes is marked as the app's. A category it sets reads "Set by an AI app you connected", not as your own choice, and the transaction's history names the app: by the name you gave its token, or for an app you signed in, by the site it comes from, such as claude.ai. It does not replace a category you chose by hand unless you ask it to, and it will not undo a split. On its own it does not teach SageFin how to categorize that merchant in future: your own choices do, and so does a batch you accept, below. It cannot write or change a note, and it cannot mark a transaction as reviewed: that says a person looked at it.

An app can rehearse a change first and see what it would do without doing it. Every change that is made is recorded in your audit log. A transaction an app adds says so at the foot of its own history, "Added to SageFin" by that app, and assigning a transaction to a business or changing its tax line shows there too.

Tags, categories and budgets

With both gates open, an app can also set up what you file transactions with, the way you would in Settings and on Budgets:

The same rules apply as when you do it. A tag name already in use is refused, and so is a budget on a transfer category or on a category whose group is budgeted as a whole. A category one of your rules sets cannot be deleted until you change or delete the rule, a group cannot be deleted while any category is in it, and a category's transactions cannot be moved onto one that already has its own budget.

Keeping a manual account up to date from its statements

Some accounts cannot be connected — an investment account at a bank no connection reaches, a mortgage, your house or your car. You can keep one current by giving its statement to an app you connected, such as Claude with the SageFin plugin, or by telling it what the thing is worth, and asking it to update the account. With both gates open, the app can:

Giving it the same statement twice changes nothing, and a statement whose dates overlap the last one adds only the lines it has not seen. The app cannot do any of this to an account a bank connection updates, because the connection is where that account's figures come from. As with every change, the app can rehearse it first and show you what would change.

None of this can move money. The write tools change how a transaction is filed inside SageFin. SageFin cannot pay, transfer or cancel anything, and a token does not change that.

When an app changes many transactions at once

A change to one transaction happens when the app makes it. A change to many arrives as a batch, and what happens to a batch is a choice you make when you create the token:

Accepting a batch teaches SageFin. For each merchant whose changes you accepted in full, the next charge from that merchant gets the same category without asking the model. A merchant with any change you unticked, or one the batch put in two categories, is not taught, and neither is one whose category you set by hand yourself. A batch applied immediately teaches nothing, because nobody looked at it.

Either way, a batch that was applied can be taken back for 30 days from the same page. Each transaction returns to the category it had, and to whoever had chosen it. A transaction you have edited since is left as it is, and you are told how many were. What the batch taught is undone too, unless you have corrected that merchant yourself since.

Only you can accept a batch. The app that proposed it has no way to, and a batch never changes a category you chose by hand or a transaction that is split.

When an app suggests a rule

An app cannot create a rule, even with both gates open. A rule acts on every transaction that arrives from then on, with nobody looking, so it is always yours to make.

What an app can do is suggest one. The suggestion appears at the top of your rules in Settings → Rules, with what it would match, what it would do, the app's reason, and how many of your transactions it matched when it was suggested. Review opens it in the rule editor, where you can change anything before saving it as a rule of your own; Dismiss removes it. Until you save it, it is not a rule and it changes nothing.

This is the same under either batch setting above. Apply immediately covers changes to transactions you already have, which you can see and take back. It never lets an app create a rule.

A suggested rule can set a category, hide what it matches, or both. Anything more is yours to add in the editor. The editor's option to also apply the rule to your existing transactions starts unticked for a suggestion.

Grant one job, not the account

A token carries only the permissions you tick when you create it, and they are narrow on purpose: a token that can read budgets cannot read transactions.

That holds for an AI app too. Connecting one needs the MCP permission, and each thing it can then ask about needs its own: account balances and the owners and beneficiaries register on If something happens need Accounts, budget figures need Budgets, spending, merchants, recurring charges, rule suggestions and your retailer orders need Transactions, and income against expenses needs Cash flow. An app asking for something its token was not given is told which permission is missing, so it can tell you, and everything else it was given keeps working. Reading your category list needs Categories. The same permissions cover the changes: creating or changing a tag needs Transactions, a category Categories, a budget Budgets, and updating an account from a statement Accounts, on top of write access.

Your retailer orders can be read, not changed. An app with Transactions can list the orders SageFin Desktop brought in from Amazon, Walmart and Costco, what was in each, and whether it was matched to a transaction, so it can answer "which Amazon orders never matched?" It cannot match, unmatch or edit an order. It sees an online order's number, as the Retail Sync panel shows it, but not an in-store receipt's. It also sees which card paid, as the panel's Paid with line does: the card's brand and its last four digits, or the last four alone for an order whose page has not named the brand, and never more of the number.

The words in your notes stay private. An app can see that a transaction has a note. It cannot read what the note says, whichever permissions its token has. SageFin's help articles need no permission beyond MCP, because they contain none of your data.

The point is what a leak costs. A token scoped to one job leaks that one job, rather than everything you have.

Signing an app in instead

Some apps can sign in to SageFin themselves, so there is no token to copy. Claude Code and the Claude desktop app are two. You add SageFin to the app, and it opens a SageFin sign-in page in your browser that asks you to approve it. Only apps SageFin has approved can do this.

An app you sign in appears under Connected apps in Settings → Integrations the first time it is used. It starts able to read only, whatever it asked for when you approved it. Can make changes lets that one app write, and the household setting above still has to be on as well.

Disconnect stops it at once. The app still thinks it is signed in, but SageFin refuses everything it asks, and signing in again does not get past that. Reconnect lets it back in, able to read only.

A signed-in app belongs to the household you connected it in. If you are removed from that household and then join another or start your own, the app does not come with you. Its next request is refused, and it appears under Connected apps in your new household as disconnected. Press Reconnect there if you want it to read this household, able to read only. Being removed also disconnects your apps in the household you were removed from.

The SageFin app on an iPhone or Android phone has the same list in the same place, so an app can be cut off from your phone without getting to a computer.

A private household

In a household in private mode, an app reads your data only for an hour you approve on your phone: an iPhone, or an Android phone that holds the household's key. The first time it asks, your phone signs it in, with Face ID (on Android, your fingerprint, face or screen lock), after showing you when the token was made or the app connected. The app can then read for up to an hour while it is working, and so can our servers, even if you sign out; it stops sooner after 15 minutes without a request. Until you approve, the app is told the household is locked and is asked to wait for you. See Private mode, under "Agents".

Revoking a token or disconnecting an app here ends its hour at once.

Tokens, and what to do if one gets out

You see the secret once, when you create it. Store it wherever the app that needs it keeps its configuration, and treat it like a password.

Give it an expiry. A year is the default. A token with no expiry is one you will forget you issued.

If it leaks, revoke it in the same screen. Revoking is immediate and does not affect your other tokens or your own sign-in, so there is no reason to hesitate — revoke first and decide afterwards whether you needed it.

Setting one up

Everything you need is in Settings → Integrations: creating the token, choosing what it can reach, and the configuration to paste into the app you are connecting, already filled in with the right address.

That screen is the reference rather than this page, deliberately — it shows the permissions that actually exist right now, and a list copied into a help article would eventually start lying to you.