What Is Claude MCP and How Does It Connect to Your Bank?

July 27, 2026 · Ledgerwire

MCP (Model Context Protocol) is Anthropic’s open standard for letting Claude read live data from outside tools, including your bank. A bank MCP connector like Ledgerwire gives Claude read-only access to your accounts:

You: What did I spend on subscriptions last month?

Claude: You had 6 recurring charges last month totaling $84.31: Netflix ($15.49), Spotify ($11.99), iCloud ($2.99), a gym membership ($39.00), and two others. Want the full list with dates?

Claude can look up balances and transactions when you ask. It cannot move money, change settings, or act on your behalf. Here’s how the bank piece actually works.

What is MCP (Model Context Protocol)?

MCP is a shared set of rules that lets an AI assistant talk to external tools in a standardized way, so every tool doesn’t need its own custom integration. Anthropic published it as an open standard: any developer can build a “server” exposing a capability (read files, query a database, look up transactions), and any MCP-compatible assistant can use it.

Before MCP, connecting an assistant to a data source meant a one-off integration built by that assistant’s own team. MCP flips that: a bank-data provider builds one server, and it works with Claude without either side writing custom code. Same idea as HTML: build to the standard once, and any compatible client can read it. That’s why “MCP” now shows up as the plumbing behind AI tools that read your calendar, your email, or your bank.

How does Claude connect to my bank through MCP?

Through a connector that bridges your bank and Claude. You never hand your bank password to Claude. With Ledgerwire, the flow is four steps:

1. You start a secure connection through Ledgerwire, which uses a bank-data aggregator (Quiltt, routing to Finicity) to link your account. You log in on your bank’s own site or app, not inside Claude.
2. Ledgerwire receives an encrypted access token from the aggregator, not your banking password, and holds it to fetch fresh data on request.
3. Claude, acting as an MCP client, sends a request through the Ledgerwire MCP server whenever you ask a money question, and gets back the live data to answer it.
4. You can revoke the connection anytime, which cuts off access immediately.

The key detail: Claude doesn’t keep a stored copy of your transactions. Each time you ask, the MCP server fetches current data for that one conversation. That on-demand model is meaningfully different from exporting a spreadsheet once and uploading it.

Is MCP safe for financial data?

MCP is just a protocol; the safety comes from how the connector implements it. We’re read-only on purpose. Three things matter most for a financial connection.

Read-only scope. A properly scoped bank MCP server exposes only read operations: balance lookups, transaction history. Ledgerwire has no “move money” code path anywhere in the connector, so there’s nothing for Claude to misuse even in principle.

Token-based access, not password sharing. Your bank credentials go directly to your bank or its aggregator during setup, never to Claude or to Ledgerwire. What we hold afterward is an encrypted access token: a revocable key, not the password itself.

You control the connection. You can disconnect anytime, which invalidates the token and ends access. Nothing about MCP requires a permanent or hard-to-remove connection.

Being direct: an MCP connection is software, and no connector can promise it’s “100% secure.” What a trustworthy one can promise is a narrow read-only scope, encrypted credentials, and a connection you can end with one click. For the deeper mechanics, see our security page.

What can Claude actually do with my bank data?

It can answer questions using your live account and transaction data, the same way it reasons over anything else in a conversation. Claude calls into the MCP server mid-conversation, fetches raw rows (merchant, amount, date), and reasons over them in plain language. No exporting, no manual categorizing. Typical questions:

• “How much did I spend on dining out in June?”
• “List every subscription I’m charged for and how much each costs.”
• “What’s my checking account balance right now?”
• “Across all my accounts, what’s my total spend by category this month?”

What are the honest limits?

Coverage depends on what your bank reports, and the data lags posting by a few days. If a field like a transaction’s merchant category isn’t reported by your bank, Ledgerwire passes it through as not reported rather than guessing or defaulting it to zero, so Claude’s answer reflects the gap instead of papering over it.

Ask about a period still in progress, like “this month,” and the answer is current-to-date, not a finished month. And because the connection runs through an aggregator, coverage and history depth vary by institution. Note too that Claude reports your actual data (what you spent, what came in, what’s recurring); it isn’t giving financial, tax, or investment advice. What you do with the numbers is your call.

What can’t Claude do with an MCP bank connection?

It can’t move money, change any account setting, or take any action at your bank, because a read-only connector has no code path for it. There’s no “pay this bill” or “transfer to savings” capability, so even if you asked, there’s nothing on the connector side to carry it out. A read-only connection can only answer questions, never act.

Contrast that with sharing a password. A login gives whoever holds it the same access you have, including whatever actions your bank’s site allows. A scoped MCP token gives Claude only the narrow read slice the connector exposes: no ambient “logged in as you” session floating around.

Do I need to be technical to set this up?

No. Connecting a bank through Ledgerwire works like signing up for any consumer app. You create an account, connect your bank through the secure linking flow (the same kind of login screen budgeting apps use), and tell Claude to use Ledgerwire. From there you ask questions in plain English: no code, no server to run. The “MCP” part is infrastructure under the conversation, the way you don’t think about HTTP when you load a webpage.

FAQ

Is MCP made by Anthropic or by Ledgerwire? MCP is an open standard published by Anthropic. Ledgerwire builds and operates the bank-data MCP server that connects to it, following the protocol’s specification.

What happens to the MCP connection if I disconnect my bank? Disconnecting through Ledgerwire revokes the access token immediately, and Claude loses the ability to query that account from that point on.

Can I connect more than one bank account through the same MCP server? Yes. Ledgerwire handles multiple linked accounts at once, so Claude can answer questions that span checking, savings, cards, or business accounts together.

Does MCP let Claude see data from before I connected? It can see whatever history your bank and aggregator make available, which for most banks is some months of prior activity, not just going forward. Exactly how far back varies by institution.

The bottom line

MCP is the open protocol that lets Claude read live data from outside tools. For banking, Ledgerwire gives Claude read-only, revocable access to your real balances and transactions, with no ability to move money, so it answers with your actual numbers instead of guesses.

Ready to try it? Connect your bank to Claude — start free, no card required.

For the protocol spec itself, see Anthropic’s Model Context Protocol documentation.