What is an MCP server? A plain explanation for your money
An MCP server gives an AI app like Claude or ChatGPT a fixed list of tools for another service. What that means for personal finance - the parts, where the data comes from, how changes work, and what to check before you connect one.
An MCP server is a program that gives an AI app a fixed list of tools it can call, using the Model Context Protocol, an open standard. A personal finance MCP server lets Claude, ChatGPT or another AI app read your own financial records, and sometimes draft changes to them, so you no longer paste numbers into a chat.
Where does MCP come from?Link to this section
- Anthropic open-sourced MCP on 25 November 2024, and called it "a new standard for connecting AI assistants to the systems where data lives".
- On 9 December 2025 MCP joined the Agentic AI Foundation, "a directed fund under the Linux Foundation, co-founded by Anthropic, Block and OpenAI".
- The official documentation names Claude, ChatGPT, Visual Studio Code and Cursor among the apps that support it.
The same documentation compares MCP to "a USB-C port for AI applications". The comparison holds up: one standard plug, so a service builds one server and every compatible AI app can use it.
What are the parts?Link to this section
The architecture overview names three:
- The host is the AI app you use, such as Claude or ChatGPT.
- The client is the connection the host keeps open to one server. A host that uses three servers has three clients.
- The server is "a program that provides context to MCP clients". For your money, that is the finance service's side.
You only ever deal with the host. The client is plumbing inside it, and the server is something you add by URL.
What does a finance MCP server give the AI?Link to this section
Mostly tools. The specification defines tools as what lets models "interact with external systems, such as querying databases, calling APIs, or performing computations". The protocol also has resources and prompts, but for a finance server, tools are where the work happens.
It works in two steps. The AI app asks the server for its list of tools, and then, while you talk, the model picks a tool from that list and the app calls it.
A real list makes that concrete. Hundo's server, as published on the setup page, has:
- 11 read tools, such as
get_net_worth,get_budget,get_iousandrun_read_query, a read-only SQL query across your own data. - 9 tools that draft a change, such as
propose_transactionandpropose_budget_upsert. Each one returns a proposal and writes nothing you own. - 1 tool that commits,
confirm_proposal, which records one proposal after you approve it.
So "the AI can see my finances" means something precise: it can call those tools, and it sees what they return. That is why the tool list tells you more about a finance server than its landing page.
Is an MCP server the same as a bank connection?Link to this section
No, and mixing the two up causes most of the confusion.
- The MCP server decides what an AI can ask for.
- The data source decides what is there to ask about. Some finance servers read accounts that a bank aggregator fills in for you. Others read a ledger you fill in yourself.
A server with a bank feed behind it knows about new card charges without you typing them. A server with no bank connection has no bank password anywhere to leak, and works in any country. Six finance MCP servers compared sets out which is which.
Hundo is the second kind. It has no bank connection at all, so what your AI reads is what you recorded: typed, imported, photographed, forwarded or dictated.
Does it run on your computer or on the internet?Link to this section
Either. The architecture overview describes two ways to connect:
- A local server runs on your own machine and talks to the AI app through standard input and output.
- A remote server runs on the provider's side and talks over Streamable HTTP.
For remote servers, the specification's authorization flow is based on OAuth 2.1. Authorization is optional in the protocol, so it is worth checking that a finance server actually uses it.
In practice, OAuth means you add a URL, the AI app opens the finance service's own sign-in page, and you approve access there.
No password and no key goes into the AI app.
Hundo's server works this way, at https://hundo.finance/mcp.
Can an MCP server change your records?Link to this section
Only if it offers tools that write, and finance servers differ a lot here. There are three designs - read-only, direct write, and confirm-first - and is it safe to connect an AI to your finances? explains what each one means for you.
The protocol itself leans towards keeping a person involved. The tools specification says there "SHOULD always be a human in the loop with the ability to deny tool invocations", and that apps should "present confirmation prompts to the user for operations".
Servers can also label their tools, for example as read-only or destructive. The same specification warns that clients "MUST consider tool annotations to be untrusted unless they come from trusted servers". A label is the server's claim, not a guarantee, so trust the server before you trust its labels.
In Hundo's case:
- The read tools are labelled read-only, and they can only read.
- The nine drafting tools write a pending proposal and nothing else.
confirm_proposalis the one tool labelled destructive, so an AI app that asks before destructive calls asks before the one call that changes your ledger.- Settings has a read-only switch. With it on, the drafting tools and
confirm_proposalare not offered to your AI at all.
How do you know a server is the real one?Link to this section
The official MCP Registry lists public servers and ties each name to a verified domain or GitHub account. It holds information about servers, not their code, and it is still in preview.
Hundo is listed there as finance.hundo/hundo, pointing at https://hundo.finance/mcp.
Two checks work for any server:
- The URL you add is on the company's own domain.
- The sign-in page that opens is the company's own, and it tells you what you are granting.
How do you connect one?Link to this section
Every remote server follows the same shape: add the URL, sign in, approve. The details differ by app:
Once it is connected, these are the questions worth asking.
Common questionsLink to this section
Do you need to be a developer to use an MCP server?Link to this section
Not for a remote one. In Claude or ChatGPT you add a URL and sign in. A local server runs on your own computer, so it has to be installed first, which is where it gets more technical.
Does an MCP server give the AI your bank password?Link to this section
Not when it uses OAuth. The AI app gets an access token for that one service, and you sign in on the service's own page. A server with no bank connection, like Hundo's, has no bank password anywhere in the system to give.
Is MCP only for Claude?Link to this section
No. It is an open standard, now a project of the Agentic AI Foundation under the Linux Foundation, and ChatGPT, Visual Studio Code and Cursor support it too.
What does Hundo's MCP server cost?Link to this section
The tracker is free forever and nothing in it is capped. The MCP server is part of Hundo Pro, which is pay what you want from $36 a year, with $72 suggested, or $8 a month.
Checked 15 September 2026 against the Model Context Protocol documentation and specification, Anthropic's announcement, the MCP blog and the MCP Registry, linked throughout. Where this page and those sources disagree, theirs win.