Most "AI for your finances" pitches start the same way: connect your bank, hand over read access to every transaction you have ever made, and let a model guess at your habits. That trade is worse than it sounds, and the output is rarely worth it.
There is a different path, and it starts from the opposite end. Track your assets deliberately. Then let the assistant you already use read the result.
A bank feed gives you thousands of rows of noise: coffee, transfers between your own accounts, refunds, duplicate pending charges. Categorising that mess is most of the work, and it is the part that breaks.
Manual tracking inverts this. You record what a holding is worth and when. Every row is one you chose to enter, in the currency you chose, at a date you decided was meaningful. Your rental property, the collectibles, the crypto in cold storage, the pension nobody's API exposes: they all sit in the same set as your brokerage account.
That is a small, honest, structured data set. It is exactly the kind of input a language model is good with, and exactly what a bank feed is not.
The Model Context Protocol (MCP) is an open standard for connecting an AI assistant to an application. It is not owned by one vendor. Claude speaks it. ChatGPT speaks it. So do Cursor, the Codex CLI, VS Code and a growing list of others.
Before MCP, "AI features" meant whatever chat box the vendor bolted onto their own product. You got their model, their prompt, their roadmap. With MCP, the application publishes what it can do, and you bring your own assistant. If you change assistants next year, the connection follows you.
For a tracking tool this fits well, because the useful questions are not the ones a dashboard answers:
A dashboard can show any one of those if someone built that screen. An assistant with access to the underlying numbers answers all of them, including the one you thought of this morning.
The tedious part of manual tracking is not the deciding. It is the typing. You log in to a broker, read six numbers off a page, switch windows, and type them in one at a time. Do that across four accounts and the monthly update stops being a five-minute job.
Modern assistants read images and files. So the update can become one message: take a screenshot of the account summary, paste it into the chat, and say "record these as today's balances". The assistant reads the numbers off the picture and writes each one. The same works for a CSV your broker exports, or a statement you drag into the window.
This is the part that only works because of the connection. Without it the assistant can read your screenshot perfectly well and then has nowhere to put the result. You would still be typing.
Three things are worth knowing before you try it:
Deposits and withdrawals work the same way, so "I put 1,000 into Degiro last Tuesday" is a complete instruction. For a large backlog, the app's CSV import is still faster, because it takes the whole file in one step.
Giving software access to financial data is where most of these stories end badly, so the boring details matter more than the demo.
Three questions are worth asking of any tool that offers this:
There is a fourth, quieter question: what happens when the assistant is wrong? Models miscount and misread. Making corrections impossible is one answer, but it is a poor one, because then every misreading stays in your history for good.
The better answer is to put a person in front of every change. An assistant can correct or remove a record here, and before it does it has to read that record back to you, show its date and amount, and get an explicit yes. One record per yes, never a batch. Your own tool asks a second time. Nothing is protected by the model choosing to be careful; it is protected by you reading the row.
None of this replaces the habit. You still decide what to track and when to update it, and that deliberate pass over your own numbers is most of the value of tracking at all.
What changes is what happens afterwards. The data set you built by hand stops being something you only read in charts, and becomes something you can interrogate in a sentence. That is a good trade, and it costs you no new credentials.
NetFiscus runs an MCP server, so you can connect Claude, ChatGPT, Cursor or any other MCP client to your own account. See the AI assistants overview, or go straight to the setup guide.
