Fio Bank

The bank I use has an API that is freely accessible to individuals (something that is not exactly common in Czech banking). The fiobank Python library conveniently wraps this API. Housekeeper pulls the transactions from the API and imports them into GnuCash, avoiding duplicates by (mis)using the transaction's unique ID we get from the bank.

In practice, an invariant that I maintain is that every split of a transaction in the "Assets:Bank:Fio" account corresponds to exactly one transaction in the bank's UI.

This is also where the bulk of the categorization happens. I find that a categorization algorithm that is effectively a bunch of if-else statements is good enough to classify something like 80% of my transactions.

if "UBER" in description:
    return "Expenses:Transport:Taxi"
elif "TESCO" in description:
    return "Expenses:Groceries"
# ...

I prefer to explicitly not categorize something that I am not 100% certain about (or possibly cannot even be certain about, such as a store that sells both groceries and medical supplies). These transactions are left uncategorized, with the other end hanging in the "Imbalance" GnuCash account until I manually categorize them.

It is worth noting that Gnucash supports arbitrary amount of splits per transaction, meaning that a single bank transaction can easily be split into multiple categories.

Of course being a fully fledged double-entry accounting software, I can even send a part of a transaction to another asset account, such as an "Assets:Money:Owed to me" account that I can use to track money that friends owe me.

Brokers

I maintain accounts with multiple brokers, and housekeeper serves to synchronize the transaction from within the broker accounts into GnuCash. One aspect I want to preserve is that housekeeper has read-only access to the broker accoutns. So, even if a broker has an API (or there exists a wrapper around the web interface API), I do not use it if it would require write access.