How to connect Claude to your bank account

For most of Claude’s short life, asking it about your money meant pasting a CSV and praying. By the time you exported, cleaned and uploaded the file, half the answers were already stale, and the categorization was whatever the model guessed from the merchant name. It worked, technically. It was also miserable.

What changed in late 2024 was the Model Context Protocol (MCP), Anthropic’s open standard that lets Claude call live tools from external apps. Combined with bank connectivity via Plaid, MCP makes it possible to give Claude a continuous, read-only view of your real finances. You ask a question in plain English; Claude calls the relevant tools; the answer is current, accurate and properly categorized.

This guide is the exact setup most people use in 2026. It takes about two minutes and works in Claude Desktop, Claude on the web (Pro/Team), and any Claude-compatible client like Cursor or Windsurf.

The architecture, in one paragraph

You don’t give Claude your bank password. You give Claude a token to a middle layer (in our case, Slate) that already has a read-only connection to your bank through Plaid. When Claude needs information, it calls a named tool on the middle layer — for example get_spending_summary or search_transactions — and gets back structured data scoped to your household. Claude never holds your bank credentials. Plaid never talks to Claude. The token between Claude and Slate is revocable in one click.

What you’ll need

• A Claude account — Claude Desktop, or Claude Pro/Team on the web
• A personal finance app that exposes an MCP server. Slate works out of the box; we’ll use it as the example below
• About two minutes

Step 1: Connect your bank (once)

Inside Slate, click Connect an account. This launches Plaid Link — the same flow you’ve seen in Venmo, Coinbase, Chime, Robinhood and roughly every fintech app of the last decade. You log into your bank inside the Plaid window, optionally pass MFA, and choose which accounts to share. Slate never sees your password. The token Plaid returns is read-only.

Transactions begin syncing within a few seconds. The initial backfill covers up to 24 months depending on the institution, which is what makes the “is this normal for me?” class of questions actually answerable.

Step 2: Generate an MCP token for Claude

Open Settings → Integrations and pick Claude. Slate generates a unique MCP URL that includes a scoped token. The token has three important properties:

• Read-only. No tool exposed through this connector can move money, change a password, or modify anything outside Slate itself.
• Scoped to your household. If you’re part of a couple or family on Slate, the token can only see your shared household — never anyone else’s.
• Time-bounded. Default expiry is 90 days; you can also pick 30 days or “until I turn it off.” A new token can be issued any time.

Issue one token per assistant. If you also use ChatGPT, generate a separate connector for it. Per-assistant tokens make revocation surgical: if you decide you’re done with one of them, you don’t lose access in the other.

Step 3: Add the connector to Claude

1. In Claude Desktop, open Settings → Connectors. On the web (Pro/Team), it’s Settings → Tools.
2. Click Add custom connector and paste the MCP URL Slate gave you.
3. Claude will fetch the tool manifest and show you the list of capabilities — things like “get accounts,” “search transactions,” “list budgets,” “get net worth breakdown.” Approve.
4. Start a new chat. The model now has access to those tools and will call them when relevant.

Step 4: Ask Claude something real

Some opening prompts that exercise the connection end-to-end:

• “What did I spend on groceries last month, and how does that compare to the same month last year?”
• “List my recurring subscriptions over the last 90 days and flag anything that looks like a price hike.”
• “Show me a chart of my spending by category for Q1.”
• “Where am I likely to go over budget this month if the rest of the month looks like the last two weeks?”
• “If I cut dining out by 30%, how much extra could I save in a year?”

You’ll see Claude call one or two tools before answering. That’s normal — well-designed finance MCP servers return narrow, semantic answers (“groceries, March, $612, vs $548 a year ago”) rather than dumping rows, so the model can reason quickly without burning context.

What to do if it doesn’t work

Claude says it doesn’t have access to that information

Usually means the connector isn’t actually attached to the current conversation. On Claude.ai, you sometimes have to toggle the connector on per chat. In Claude Desktop, quit and reopen after adding a new connector.

Some categories look weird

Categorization is a long tail. Open Slate, find the transaction, and set a merchant override. Future syncs — and every future Claude answer — will reflect it.

You want to disconnect

Settings → Integrations → Revoke. Access cuts off immediately on Claude’s next tool call. You can also revoke Plaid’s access to your bank from inside your bank’s own app for full belt-and-suspenders.

Is this safe?

Done through a proper middle layer, yes. The threat model worth thinking about is narrower than people expect: Claude can read your money, not move it. The worst-case outcome of a compromised Claude session, for a properly-scoped MCP token, is that someone learns what you spent on groceries — not that they drain your account. That’s a very different risk surface than e.g. handing out your bank login.

The non-negotiables for any setup you trust: read-only by default, per-user tokens, an honest privacy policy that says outright the provider doesn’t sell your data or train models on it, and one-click revocation. Slate is built on all four.