Resources
How to connect bank data to an AI-built budgeting app in Australia
Build an Australian expense tracker with Fiskil. Follow one account from consent to a transaction list, posted spending totals and categories your users can edit.
A budgeting app can show the wrong spending total even when it has the right bank transactions. Take a market shop for A$62.40, breakfast for A$8.60 and a pending rail fare for A$5.20. All three belong in the list. Only the first two belong in today’s posted spending total of A$71.00.
Here’s how to build that screen for an Australian user with Fiskil. Your backend connects the bank with the user’s consent and fetches the transactions. Your app decides what counts as spending and remembers the categories the user chooses. The amounts and labels below are a fictional example you can use to test your own app.
1. Connect your coding assistant and your backend
If you’re using an AI coding assistant, give it Fiskil’s documentation MCP address, https://docs.fiskil.com/api/ai/mcp. It can look up endpoint definitions and examples without a Fiskil API key. Your application needs its own credentials to fetch bank data. Keep those in your backend’s secrets store and use placeholders in prompts. Fiskil MCP setup for supported clients.
In Claude Code, run claude mcp add --transport http --scope project fiskil https://docs.fiskil.com/api/ai/mcp, then check the connection with /mcp. In another assistant that supports remote MCP, use its HTTP connector settings. If you’re working without MCP, give the assistant the endpoint links below. Fiskil’s client setup guide.
Give the assistant a concrete job.
Add a Connect your bank button and a transaction list to my budgeting app. Use Fiskil’s sandbox. Look up the current End User, auth session and banking endpoints. Keep pending purchases visible, calculate monthly totals from posted purchases, and preserve my category edits when the bank feed refreshes.
Create sandbox credentials in the Fiskil Console. Your server exchanges client_id and client_secret at POST /v1/token and sends the resulting bearer token on API calls. This example selects v3 with X-Fiskil-Version: v3. The banking request paths below still begin with /v1/. Authentication · API versioning.
2. Take one user from consent to their first account
When Alex signs in, create a Fiskil End User and save the returned end_user_id against Alex’s app user ID. When Alex taps Connect your bank, look up that mapping on the server. Create an auth session for that End User with your success and cancellation URLs. Open the returned auth_url, or integrate the Link SDK. Alex selects a bank and the accounts to share, then approves the request. Create End User · Create auth session · Link SDK.
On return, show “Loading transactions”. The consent.received webhook confirms consent. The separate banking.transactions.sync.completed webhook tells your backend when to fetch transactions. Verify the signature against the original request bytes and make it safe to receive the same event twice. Look up the accounts available to Alex using GET /v1/banking/accounts with the mapped end_user_id. Check the account belongs to Alex before serving a budget or accepting an edit. Webhooks · Accounts reference.
3. Keep the bank record and the user’s edit in separate fields
These are suggested records in your app’s database. The names in the left column are application choices. Fiskil’s references describe the account and transaction data to map into them. Accounts · Transactions.
| App record | What to keep |
|---|---|
| User and connection | App user ID, Fiskil End User ID, consent ID and sharing state. Check active access for each account. |
| Account | Owner, institution, account ID and display name. Use those identifiers even if two accounts have the same name. |
| Bank transaction | Owner, account, Fiskil transaction ID (fiskil_id), status, signed amount, currency, posting and execution times, description and source category when available. |
| Budget edit | Owner, account, transaction ID and the chosen budget category. Keep this record when refreshing bank fields. |
| Import progress | Owner, account, query window, next page and last completed import. Keep the previous complete view visible while a new import is in progress. |
In this example the app stores AUD amounts as integer cents, so a debit of A$8.60 becomes -860. The budget uses the Australia/Sydney calendar month. Keep other currencies in separate totals.
4. Fetch the transaction list, including page two
Suppose Alex selects their everyday account. From the server, try a small page size so you can exercise the second page. Substitute the End User and account identifiers you resolved for Alex. Leave the status filter unset to retrieve both pending and posted rows. Get Transactions.
curl --get 'https://api.fiskil.com/v1/banking/transactions' \
-H "Authorization: Bearer ${FISKIL_ACCESS_TOKEN}" \
-H 'X-Fiskil-Version: v3' \
--data-urlencode "end_user_id=${END_USER_ID}" \
--data-urlencode "account_id=${ACCOUNT_ID}" \
--data-urlencode 'page[size]=2'
Read the transactions array. Insert or update each row using the owner, account and fiskil_id as the key. Follow the response’s links.next until there is no next page. Save the page’s rows and its checkpoint together so you can retry a failed page. Mark the import complete only after the last page. The same logic applies when a repeated webhook starts another read. Transaction parameters · Fiskil’s worked pagination and record-ID example.
For date-filtered requests, Fiskil checks the posting time first, then the execution time. Rows with neither time fall outside a date-filtered result. Use the same order when assigning a transaction to the user’s reporting month. Show an undated item separately if you receive one in an unfiltered list. Banking date behaviour.
If your screen also shows the account balance, fetch it from the balances endpoint. An account balance includes movements such as wages and transfers, so it answers a different question from a spending total. For a “last updated” label, check freshness for banking.transactions. A transaction refresh queues work asynchronously. Keep the prior complete view visible until fresh rows arrive. Get Balances · Get Data Freshness · Refresh Banking Transactions.
5. Turn the list into a spending total
Here’s Alex’s November test feed. The app knows the savings transfer is between Alex’s own accounts. The budget measures gross posted purchases. A different metric, such as spending after refunds, needs its own rule.
| November activity | Signed AUD | Budget treatment |
|---|---|---|
| Market shop | −62.40 | Posted purchase · Groceries |
| Station bakery | −8.60 | Posted purchase · initially Groceries |
| Rail fare | −5.20 | Pending · show separately |
| Salary | +900.00 | Posted income · outside spending |
| Transfer to own savings | −100.00 | Posted transfer · outside spending |
Sum the magnitudes of posted purchase debits for this owner, account, month and currency. The market and bakery give A$71.00 in Groceries. The pending fare adds A$0.00 to that figure. Calling every negative amount a purchase would also put Alex’s A$100 savings transfer into the budget. In your app, let users review transfers you can’t confidently identify. Fiskil’s signed-amount and posted-status example.
Say the fare later posts at A$4.80. After refreshing the list, count the posted fare once. The total is now A$75.80. A matching description or amount is too weak to prove that a pending and posted row represent the same payment. Use a confirmed record relationship where available. Otherwise refresh the pending view from the latest complete result for the same scope and keep unresolved pending items outside the total. Avoid moving a user edit to a different transaction by guessing. Transaction status filter · Data freshness.
6. Let Alex change breakfast to “Eating out”
Alex taps Station bakery, chooses Eating out and saves. Authorise the edit against Alex’s account, then upsert the choice against that transaction. When building the chart, use the user’s category first, then the mapped feed category, then Uncategorised. A refresh can update the bank’s description or category while preserving Alex’s choice.
Before the fare posts, the edited chart shows Groceries A$62.40 and Eating out A$8.60. Total spending stays at A$71.00. After it posts, add Transport A$4.80. Let Alex clear the override to use the current feed category again.
This query runs against your own tables. Here, budget_month is the local reporting month, kind is the app’s classification and budget_category is the feed category mapped to a budget label. Enforce one row per owner, account and transaction in each table so an edit can’t multiply a purchase in the join.
SELECT COALESCE(e.category, t.budget_category, 'Uncategorised') AS category,
SUM(-t.amount_cents) AS spent_cents
FROM bank_transaction AS t
LEFT JOIN budget_edit AS e
ON e.owner_id = t.owner_id AND e.account_id = t.account_id
AND e.fiskil_id = t.fiskil_id
WHERE t.owner_id = :owner AND t.account_id = :account
AND t.budget_month = :month AND t.currency = 'AUD'
AND t.status = 'POSTED' AND t.kind = 'purchase' AND t.amount_cents < 0
GROUP BY COALESCE(e.category, t.budget_category, 'Uncategorised');
Personal labels such as “Friday breakfast” can live entirely in your app. If your integration also needs to change Fiskil’s predicted category, its category endpoint takes fiskil_transaction_id (the transaction’s fiskil_id) and a valid secondary_category from Fiskil’s taxonomy. Keep a free-text budget label local. Override Transaction Category · How to evaluate transaction categories.
7. Test the whole screen before requesting live access
Start in the Fiskil sandbox with a test End User, select the sandbox bank, complete the verification step and share an account. Use that connection for the consent and API checks. Use the fictional rows above as a repeatable app test for exact amounts, edits and posting transitions. Sandbox walkthrough · Australian banking API proof-of-concept guide.
- Cancel the bank connection, then complete a fresh attempt. Check that the app shows the difference between cancelled, syncing and a completed import with zero rows. Auth sessions.
- Interrupt page two, retry it and replay the sync event. Check that the completed list has each purchase once and the category edit survives. Transactions.
- Replay the rail fare first as pending, then as posted at the lower amount. Check the two totals against A$71.00 and A$75.80.
- Sign in as a second test user and try to read or edit Alex’s transaction. Both operations should fail. End Alex’s consent and check that affected imports stop and the app offers reconnection. Consent lifecycle.
Once the sandbox screen behaves as expected, request live access through the Fiskil Console. Use the production checklist to review your company profile, security questionnaire and final end-to-end test before connecting real customers. Go-live checklist · Open the Fiskil Console.


