A scope is the list of things an AI connector is allowed to do in your finance app. Least privilege means giving it only what the job needs: reading for questions, a narrow write permission for a specific clean-up, and nothing else. In SpendixAI you choose the scope when you connect Claude or ChatGPT, can start read-only, and can disconnect at any time.
What is a scope, exactly?
When an assistant connects, the app asks you to approve access. The approval is not all or nothing. It comes in pieces, each a permission such as reading transactions or changing them. The assistant receives a token that carries exactly the permissions you approved, and the app checks that token on every request. A request outside the scope is refused, however politely the assistant asks.
Why start read-only?
Reading cannot damage your records. The worst case is that the assistant learns something you would have preferred to keep private, and that is limited by what you chose to expose. Writing can change categories, create transactions or remove them. Even a well-meant misunderstanding becomes a clean-up job.
A way to think about scope levels
The labels on your connection screen are what count, so read them. As a general pattern, expect tiers like these.
| Level | What it allows | When it fits |
|---|---|---|
| Read-only | Looking at transactions, categories and reports | Questions, summaries, spotting odd rows |
| Read plus limited changes | Also adjusting categories or rules | A one-off clean-up of history |
| Read plus adding records | Also creating new transactions | Entering cash spending in batches |
| Broader access | Anything the connection exposes | Rarely needed; avoid by default |
Worked example: matching tasks to scopes
Say you have four jobs this month. Job 1 is asking what you spent on groceries: read-only. Job 2 is fixing 12 rides in the wrong category: a scope that allows category changes. Job 3 is adding 8 cash entries: a scope that allows creating records. Job 4 is asking about fuel again: read-only. Without least privilege you would hold the broadest scope for the whole month, 4 jobs out of 4. With it, you hold a broad scope for 2 of 4 jobs, which is half the exposure, and you drop back to read-only between them.
How do you audit a connection?
- List every assistant currently connected.
- For each, read the scope it holds.
- Ask whether you used it in the last month.
- Disconnect any you did not use.
- Reduce the scope on those you did, to what you actually need.
What should developers look for?
- Whether scopes are per action or per area of data.
- Whether tokens expire and how renewal works.
- Whether there is a log showing what the assistant did.
- Whether a write scope can be limited to a category or account.
- Whether the app asks again if the requested scope changes.
Treat anything not documented as absent. If the connection screen does not show it, do not assume it exists.
What about tokens and revoking access?
Behind the screen, an approved connection holds a token. Treat it like a key: you never need to see it, copy it or paste it anywhere. If you ever suspect an assistant account was compromised, disconnect the connection from the app side first, then secure the assistant account. Reconnecting afterwards creates a fresh approval, with a scope you can set again from scratch. That is the practical value of least privilege: when something goes wrong, the damage is limited to what was allowed.
What scope does not cover
A scope limits what the connector can do inside the app. It does not control what the assistant provider does with the words you and the assistant exchange. Read that provider's terms, keep identifiers out of your messages, and remember that SpendixAI has no bank connection and never asks for net-banking logins, so the assistant cannot reach your bank through it.
Useful ideas start with a clearer picture of your own spending.
Explore your reports


