BankMCP™

BankMCP™

Your AI now reads your bank. Ask it anything about your accounts. Read-only, self-hosted, one user. Standard MCP; tested with Claude and Ollama.

BankMCP™ is not a bank. It is a small open-source server you host yourself (npm package bankmcp). It connects to your banks through Enable Banking, which wraps 2,700+ European banks in one PSD2 API, and exposes them to any MCP client as a connector. Read-only, no payments, no third party holding your data.

"Has the invoice from Acme been paid?" · "What did we spend on groceries in August?" · "Which subscriptions am I paying for, and what do they cost per year?" · "Tell me when my balance drops below 5,000."

How it works

Your assistant ──OAuth──▶ your BankMCP™ server ──JWT──▶ Enable Banking ──PSD2──▶ your bank

• Your assistant (Claude, ChatGPT, Cursor, or any MCP client) talks to your server as a connector. You sign in once with a password; tokens handle the rest.
• Your server holds the Enable Banking application key, the bank consents and your account ids. It does not store balances or transactions and sends no telemetry.
• Enable Banking is the licensed provider between your server and your bank. Every balance and transaction you ask for passes through their servers on the way to yours; they do not store it, and they do not see your bank credentials, since you log in at your bank's own site. This hop is how PSD2 works and is the one part of the chain that is not on your machine.

Setup

Two ways to run it. Both need a free Enable Banking account and about ten minutes.

• On your own machine for Claude Desktop, Claude Code, Cursor and other desktop MCP clients. Nothing to deploy, no password.
• On a small server when you want it in claude.ai or on your phone.

On your own machine

The server and its state live on your computer, and no AI vendor or app maker sees your data. The bank connection is not on your computer: it goes through Enable Banking, as described above.

Requires Node 24 or newer. Add BankMCP™ to your client:

Claude Code:

claude mcp add bankmcp -- npx -y bankmcp

Claude Desktop: download bankmcp.mcpb and open it; Claude Desktop installs it as an extension. Or add it by hand in claude_desktop_config.json:

{ "mcpServers": { "bankmcp": { "command": "npx", "args": ["-y", "bankmcp"] } } }

Cursor and others: the same command, npx -y bankmcp, as a stdio server.

Then ask your assistant anything about your bank. It will answer with a localhost address. Open it: the setup page lists the values to register an application at Enable Banking, then takes the application id and the key file. Your browser will warn once about the certificate on localhost, which the server made for itself because Enable Banking requires https for the bank redirect. Continue past it. Say "connect my bank" and log in at your bank.

To avoid the warning altogether, use a certificate your browser already trusts: mkcert makes one with mkcert localhost 127.0.0.1; point TLS_CERT_PATH and TLS_KEY_PATH at the two files and the server uses them instead of generating its own. BankMCP™ itself does not touch your system's trust store.

State lives in ~/.bankmcp. Delete the folder to forget everything.

On a server

1. Deploy

Any container host works. The server needs a persistent volume at /data and a public https address; it asks you for everything else in the browser. A prebuilt image is published on every release: ghcr.io/noskillish/bankmcp (built by GitHub Actions from the tagged source, linux/amd64 and linux/arm64).

Railway: [Deploy on Railway] deploys the published image with a volume at /data and a domain, nothing to fill in. Or by hand: New Project, Deploy from GitHub repo, pick this repo, add a volume mounted at /data and generate a domain (Settings, Networking, port 8080). The server learns its own address from Railway.

Fly.io: three commands from a checkout, using fly.toml:

fly launch --copy-config --no-deploy   # pick an app name and regionfly volumes create data --size 1fly deploy

Render: [Deploy to Render] uses render.yaml. Render's persistent disks need a paid instance.

Docker on your own box:

docker run -d --name bankmcp --restart unless-stopped \  -p 8080:8080 -v bankmcp-data:/data ghcr.io/noskillish/bankmcp:latest

or docker compose up -d with compose.yaml. Then put a TLS terminator in front (Caddy needs two lines: YOUR-HOST { reverse_proxy localhost:8080 }) and set BASE_URL to the public address. The server binds every interface by default. When the proxy runs on the same machine, set BIND_HOST=127.0.0.1 so nothing else on the local network reaches the login and setup pages.

Open the address. A fresh server shows a setup page.

2. Register an Enable Banking application

The setup page can do this for you. Enter the email of your Enable Banking account (or the one you want to create it with) and press Send me the sign-in link. Enable Banking emails you a one-time link; paste it on the next page. Your server signs in with it once, creates the application with the right redirect and policy addresses, generates the key pair and stores the id and the key. The sign-in is not kept. On a local server the click on the link does the same without pasting.

Or by hand: the setup page lists the exact values Enable Banking's form asks for, with Copy buttons. At https://enablebanking.com/cp/applications create an application with them, keep generate private key selected, and save. A .pem file downloads once; that is the key. Paste the application id and choose the .pem file on the setup page. Before anything is stored, the id and key are checked against Enable Banking.

Either way: Production for your real accounts, Sandbox for test data. A production application for your own accounts is activated once in the Control Panel with Activate by linking accounts; see Going live below.

3. Finish setup

Choose a password of twelve characters or more. A Ready page then shows what was checked: the application, whether its redirect URL is registered and whether it is active, and the connector address to copy.

Prefer configuration by environment? Set these and the setup page does not appear:

4. Add the connector in your assistant

In claude.ai (or the desktop app): Settings → Connectors → Add custom connector. Name it BankMCP™, paste https://YOUR-HOST/mcp, save, then click Connect. Your server shows a password page; enter the admin password. That is the only login you will do.

In Claude Code:

claude mcp add --transport http bank https://YOUR-HOST/mcp

then run /mcp inside Claude Code to sign in.

Other MCP clients (ChatGPT, Mistral Le Chat, Cursor, VS Code) work the same way: add the URL as a remote MCP server, sign in with the password. Tested with Claude, Claude Code and Ollama; the others follow the same standard. A client whose domain is not in ALLOWED_REDIRECT_HOSTS needs adding there.

5. Connect your bank

In your assistant, say "connect my bank" (or use the connect-bank prompt). It looks up your bank, gives you a link, you log in at the bank and approve, and the accounts appear. Consents last up to 180 days; you are told when one is about to expire and the same conversation renews it.

Give accounts labels ("Everyday", "Joint expenses", "Mortgage") when it suggests them. Every tool accepts labels instead of ids.

Going live with your own accounts

Enable Banking's production environment normally requires a contract, but it has a restricted mode for accessing your own accounts, explicitly allowed for individual non-commercial use. After registering a Production application:

1. On the (Inactive) application click Activate by linking accounts.
2. Log in at your bank and approve. Repeat for each bank you want.
3. The application becomes active and the API returns only the accounts you linked this way.

Read the Restriction of Use section of Enable Banking's Terms of Service before you rely on it: restricted mode is for your own accounts, not for offering a service to others. This project does not change those terms.

What you get

Tools (all read-only):

Prompts: connect-bank, monthly-summary, build-budget, savings-scan, subscription-audit, unusual-transactions.

Watches run on the server. Rules: balance below or above an amount, a single debit over an amount, an incoming or outgoing payment matching a name, and "tell me if this payment has not arrived by this date". Accounts are checked at most four times a day, the PSD2 limit for unattended access. Notifications go to NOTIFY_WEBHOOK_URL as a Slack message or a JSON POST.

Enable Banking's own webhooks cover payment initiation only, so account data is polled. There is no way around that under PSD2.

Claude Code plugin

The repository is also a Claude Code plugin marketplace. The bank plugin brings three skills: /bank:setup walks you through installing BankMCP™ on your machine, /bank:deploy through hosting it, and bank encodes how to work with the data: an account map, categorisation rules, the monthly review format and when to create watches.

Point it at your server, then install:

export OPENBANK_URL=https://YOUR-HOST/mcp   # put this in your shell profile

/plugin marketplace add noskillish/bankmcp/plugin install bank@bank

Then /mcp, select bank, Authenticate, and enter your password. No organisation admin is involved; plugins are per user.

The skill lives at plugin/skills/bank/SKILL.md. Copy it into your own skills to fill in the account map and your merchant rules. The server stays generic; your rules stay yours.

Local models (experimental)

The server does not care which model asks. npm run chat bridges an Ollama model to the same tools over stdio, so no AI vendor sees a transaction:

ollama pull qwen3:8bnpm run chat -- "what's my balance?"

Measured on a MacBook Air with 24 GB: correct per-account balances, a wrong total, ten minutes per answer. An 8B model is not yet trustworthy with money; a 30B-class model on a machine with a real GPU is where it gets useful. Any MCP client with tool calling (LM Studio, Goose, Jan) can also point at node src/stdio.ts directly. Enable Banking still sits between you and the bank either way; that part is regulated and unavoidable.

Security notes

• The server is a complete OAuth 2.1 authorization server with one user. Discovery, dynamic client registration and PKCE come from the MCP SDK; tokens are stored hashed; five wrong passwords lock an address out for fifteen minutes, and twenty from all addresses together lock sign-in for everyone for fifteen minutes. Assistants that are already connected keep working during a lock.
• State is one JSON file in DATA_DIR: consents, account ids, watches (with the ids of transactions that already fired) and OAuth tokens. Balances and transactions are not written to disk. Your assistant keeps the conversation as any chat does, and a watch notification carries the matched transaction to your webhook. Back the file up if you care about not re-consenting; delete it to forget everything.
• Only clients that redirect back to a known MCP client domain (Claude, ChatGPT, Mistral, Cursor, VS Code) or localhost can register (ALLOWED_REDIRECT_HOSTS), which stops a phishing link from routing your sign-in to another site. Using a client not on the list? Add its domain. The sign-in page also names the host you will be sent back to.
• Notifications go to one destination only: the NOTIFY_WEBHOOK_URL the operator sets on the host, and it must be https. A watch cannot name its own webhook, so a client or an assistant that has been talked into it cannot point transaction data at another address or at something inside the server's network.
• Where the checks are, for anyone reading the code: every /mcp request passes requireBearerAuth in src/app.ts before any tool runs; tokens are issued and verified by SingleUserProvider in src/auth.ts, after the password check in completeLogin. There is one user, so a valid token is the whole authorisation and the tools carry no further per-user checks. In local mode (npx bankmcp) there is no OAuth: the client on your machine talks to the server over stdio, and whoever can run processes on that machine can use it.
• Transaction descriptions and counterparty names are written by other people, for example whoever sent you a payment. The server tells assistants to treat them as data, not as instructions. An assistant can still be misled by them.
• Anyone with the admin password can read your accounts. Use a long one. Every successful sign-in is logged and, if NOTIFY_WEBHOOK_URL is set, sent to you as a message. A sign-in you did not make is your alarm.
• Forgot the password? Set ADMIN_PASSWORD (or ADMIN_PASSWORD_HASH) on the host and restart; the environment wins over the value stored at setup.
• Changing ADMIN_PASSWORD_HASH (or ADMIN_PASSWORD) and restarting logs every client out. That is the kill switch. Revoking the consents at your bank, or deleting the state file, is the step beyond it.
• There are no payment tools. Payments need a licensed PISP and a very different security model, so they are out of scope.

Commands

npm start              # http server (reads env from the environment)npm run dev            # same, with reload and .envnpm run check          # verify config and the Enable Banking applicationnpm run hash-password  # produce ADMIN_PASSWORD_HASHnpm run watch -- --force   # run all watches once, print what firednpm test               # unit tests (node:test)npm run typechecksh scripts/build-mcpb.sh   # Claude Desktop bundle → dist/bankmcp.mcpbnpm run banks          # refresh docs/banks/banks.json from Enable Banking's bank listnpm run site           # build the subpages of the landing page into docs/

Requires Node 24 or newer (runs TypeScript directly, no build step).

Layout

src/stdio.ts          entry point for npx bankmcp: stdio server plus the localhost https sidesrc/server.ts         entry point for a hosted deployment: OAuth, /mcp, callback, status pagesrc/app.ts            the Express app shared by both entry pointssrc/local.ts          localhost https server and self-signed certificate for local modesrc/mcp.ts            McpServer factory (tools, prompts, instructions)src/tools.ts          the MCP toolssrc/prompts.ts        the MCP promptssrc/watcher.ts        background rule checks and notificationssrc/auth.ts           single-user OAuth providersrc/setup.ts          first-run setup: validates and stores id, key and passwordsrc/pages.ts          the HTML pages: setup, sign-in, connected, status, privacy, termssrc/config.ts         environment and data-directory configurationsrc/enablebanking.ts  JWT signing and a thin typed API clientsrc/store.ts          the JSON state filesrc/data.ts           shaping balances and transactions for an assistantsrc/cli.ts            check, hash-password, watchbin/bankmcp.js        the npx entry; dispatches to stdio or cliscripts/              build-mcpb.sh (Claude Desktop bundle), banks.ts and site.ts (landing subpages), local-chat.ts (Ollama bridge)plugin/               Claude Code plugin: setup and deploy skills, bank skilldocs/                 landing page (GitHub Pages, bankmcp.dk); docs/banks holds the bank list and its pages

What this is, and is not

BankMCP™ is software, not a service. There is no hosted BankMCP™, no account to sign up for, and nobody but you handles your server, your key or your bank consents. Each person who uses it deploys their own copy and is the sole operator of that copy: they register their own Enable Banking application, accept Enable Banking's terms themselves, and are responsible for their own hosting, password and security.

The authors publish the code and nothing else. They do not run any instance for others, receive no data, and are not affiliated with Enable Banking, Anthropic or any bank. BankMCP™ is not a bank, does not hold money, and gives no financial advice.

Attestations (ATTESTATION_ENABLED, off by default) do not change any of this. They add no payment initiation and widen no consent: the bank calls underneath are the same balance reads get_balances already makes, and a third party who asks learns a yes or a no about one IBAN they already knew. Whether sharing that with a particular service suits your Enable Banking terms is your decision as the operator, as everything else here is.

Use at your own risk. If you deploy it, you own that deployment and its security. The software is provided as is, without warranty of any kind, and the authors accept no liability for how it is used or for any loss that follows. MIT licence below.

License

MIT