You can analyse your spending with Claude without a CSV export by connecting it to your expense tracker through MCP, so it reads categorised data on request. Start read-only, ask a precise question, and verify the number against the tracker.
If you build small tools for yourself, you will recognise the appeal: no export step, no stale file, no parsing script to maintain.
What changes when you drop the CSV?
- The data is live at the moment you ask, not as of the last export.
- You stop maintaining a parse-and-clean script for every statement layout.
- Categories already applied in the tracker come along, so Claude works on tidy data.
- You need to decide scope, since the assistant can now reach the source.
SpendixAI supports connecting Claude or ChatGPT through MCP. You choose a scope, can start read-only and can disconnect any time.
Good prompts for a connected Claude
- How did my spending by category change between September and October?
- List my five largest transactions this quarter with merchant and category.
- Which merchants appear more than three times in a month?
- Which transactions have no category yet?
- Draft a monthly budget from the last three months of actuals, and show your working.
Ask it to show its working. That makes mistakes visible and lets you spot-check one line.
A worked check
Say your tracker shows October spend of ₹41,200 and September of ₹36,800. The change is 41,200 minus 36,800 = ₹4,400, which is 4,400 divided by 36,800, about 12 percent. If Claude reports a rise of ₹4,400 and roughly 12 percent, it matches. If it says 21 percent, you have found a slip worth one correction.
| Measure | Tracker | Claude said | Check |
|---|---|---|---|
| September total | ₹36,800 | ₹36,800 | match |
| October total | ₹41,200 | ₹41,200 | match |
| Rise | ₹4,400 | ₹4,400 | match |
| Rise in percent | about 12 | about 12 | match |
How do I structure a prompt for repeatable analysis?
Give Claude the period, the grouping and the output shape, and ask for the working. For example: for October, group spending by category, list the top five by amount, show each as rupees and as a share of the total, and show how you added them. A repeatable prompt works like a saved query; you can reuse it every month and compare the outputs.
- State the period explicitly, such as 1 to 31 October.
- Say which categories to include or leave out.
- Ask for amounts in rupees and for the arithmetic.
- Ask it to list any rows it could not categorise.
How do I turn the analysis into something I keep?
Copy the verified figures into your own notes or a sheet, with the month beside them. A rolling log of monthly totals by category lets you spot drift over a year with no export at all. Because the numbers come from a tracker that holds each transaction with its account and category, you can always drill back to the rows behind any figure.
If you keep a personal script, let it read the figures you have verified rather than the raw chat output, so the checking step stays in the loop. A small habit that pays off is to save each month's prompt text beside its verified answer, so next month's comparison uses exactly the same wording and any difference comes from the data, not the phrasing.
What it cannot do
- It cannot see spending that was never recorded; cash must be typed in.
- There is no bank sync or SMS reading, so statements come by upload.
- It will not fix miscategorised rows on its own; correct them in the tracker so corrections become history.
- It does not remove the need to read the assistant provider's data terms.
Useful ideas start with a clearer picture of your own spending.
Explore your reports

