Safe enough depends on what you allow
Nobody can honestly promise that connecting an assistant is safe in every case. What you can do is limit what it may touch, keep a person in the loop, and know how to leave. The checklist below comes from how SpendixAI, our own expense tracker for India, handles MCP connections, plus what the assistants' documentation says about approvals. Work through it once before you connect, and again when you change anything.
The checklist
- Use the standard sign-in. SpendixAI uses OAuth with a consent screen, so there is no API key or secret to paste into a file. If a guide tells you to paste one, stop.
- Choose the workspace deliberately. The consent screen asks which workspace to share, and a connection covers only that one.
- Pick Read only first. Move up to Read and add, or Full access, only for a specific task.
- Check what your assistant asks before it acts. Look at its approval setting and turn on prompts, particularly for tools that add or change data.
- Remember who writes the answers. The assistant's provider produces them, under its own terms.
- Verify figures that matter. For tax numbers especially, compare with the matching SpendixAI report.
- Review the connection list now and then, and disconnect anything you no longer use.
Where each check happens
| Check | Where you do it | What to look for |
|---|---|---|
| Preset | SpendixAI consent screen | Read only selected |
| Workspace | SpendixAI consent screen | The right workspace chosen |
| Approvals | Inside the assistant | A prompt before tools run |
| Connected assistants | Developer → MCP connections | Only ones you recognise |
| Figures | SpendixAI reports | Same totals as the assistant said |
Approvals differ by assistant
Claude shows per-tool approval prompts, offers an Allow always option that its documentation says to use only for trusted servers, and lets you set individual tools to Blocked. Cursor asks for approval before using MCP tools by default. Cline treats an empty autoApprove list as asking every time. Zed's default is to confirm. Codex can be set to prompt for tools not marked read-only. Mistral Vibe asks for each function unless you pre-authorise it. ChatGPT asks before write actions by default. goose is the one to watch, because its documentation says it operates autonomously by default.
What SpendixAI itself limits
Even at the widest preset, billing and member settings are not available to an assistant. Your workspace role is a second ceiling, and access is checked on each call. SpendixAI has no bank connection and does not read SMS, so an assistant cannot reach beyond the records you have already put in your workspace.
What SpendixAI cannot control
Once data is shown to an assistant, how its provider handles it is governed by that provider's terms. Anthropic's documentation also notes that custom connectors are not verified by Anthropic, which is a good reminder that you are the one vouching for a connection. If that is uncomfortable, stay on Read only, or skip assistants for the questions that feel too sensitive.
A five-minute test run
Before you rely on a new connection, try it with questions whose answers you already know. Ask for last month's total in one category and compare it with the report. Then ask something the assistant should refuse or be unable to do, such as changing your billing details, and confirm it cannot. If the assistant asks permission before using a tool, read the request instead of clicking through it. The goal is to learn how the assistant behaves while the stakes are low.
Note the result for each assistant you try. A short list of which ones asked first and which ones did not will tell you more than any general advice.
If something looks wrong
Disconnect first, ask questions later. Developer → MCP connections lets you cut an assistant off at any time. Then look at your transactions in SpendixAI. Deleted transactions go to a recycle bin for 30 days, which helps if a write action removed something you needed.
Useful ideas start with a clearer picture of your own spending.
Explore your reports
