AI NewsDev toolsAnnouncement

BankMCP lets AI assistants read your own bank accounts through MCP

A one-person MIT-licensed project called BankMCP turns Enable Banking's PSD2 API into a self-hosted MCP server, so an MCP client such as Claude Code, Cursor, ChatGPT or Ollama can read balances and transactions from 2,700 European banks with no third party holding the data.

AI News

Editorial2 min read

LinkedInX
BankMCP repository page on GitHub

Why it mattersA developer can connect personal bank data to their coding-agent workflow the same way they connect a database or an API, without giving an AI vendor or a financial aggregator control of the credentials or the transactions.

BankMCP is a self-hosted, read-only Model Context Protocol server that sits between an MCP client and a European bank. The project was released under MIT on 7 September and is at 159 stars on 8 September, per the GitHub REST API. The author, GitHub user noskillish, calls it "software, not a service": every user runs their own copy against their own accounts.

How the parts connect

The server connects to banks through Enable Banking, a licensed PSD2 provider that gives access to 2,700 European banks through one API. BankMCP exposes that connection as an MCP server, so any client that supports the standard, including Claude Desktop, Claude Code, Cursor, VS Code, Mistral Le Chat and ChatGPT, can call tools such as list_accounts, get_balances and get_transactions. The README says the server has been tested with Claude, Claude Code and Ollama.

The BankMCP host holds the Enable Banking application key, the consents and the account ids, and never writes balances or transactions to disk. The README calls Enable Banking the one part of the setup that is not on your machine, which is what lets the project cover any bank Enable Banking supports without a custom-built PSD2 client for each one.

Restricted mode is what makes this usable

Enable Banking's production tier normally needs a commercial contract. BankMCP runs on Enable Banking's "restricted mode", documented as available for individual, non-commercial access to your own accounts: you activate the application by linking each of your own banks, and the API returns only those accounts. The README tells readers to check Enable Banking's Terms of Service and states the project does not change them.

There are no payment tools. The README says payments need a licensed Payment Initiation Service Provider and a different security model, so they are out of scope. The set of tools stays small: connect and manage banks, list accounts, get balances, paginated transactions, plus a background watch system with webhook notifications for rules such as "balance below an amount" or "an incoming payment matching a name".

The security notes are the reason to install it

The server is a single-user OAuth 2.1 authorization server, tokens are stored hashed, and five wrong passwords lock an address out for 15 minutes. State is one JSON file in the data directory; deleting it forgets everything. Changing the admin password and restarting logs every client out.

The author is also direct about local models. On a 24 GB MacBook Air, the README reports that Ollama with qwen3:8b returns the right per-account balances but a wrong total, and takes ten minutes per answer. The stated conclusion: an 8B model is not yet trustworthy with money and a 30B-class model on a real GPU is where it becomes useful. Anyone connecting bank data to an agent workflow can now install the connector, keep the transactions on their own computer, and pick which MCP client and which model they trust with the questions.

Source

SourceGitHub

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project