About This Project

Bunq Atlas - Hackathon Project

Bunq Atlas is an intelligent finance assistant that combines the power of Claude AI with the bunq banking API through Model Context Protocol (MCP).

Inspiration

This project was born out of pure frustration with existing AI assistants in the banking space. Our team encountered "Finn" - an AI assistant so incompetent it couldn't even handle a simple refund request. While Grok AI and ChatGPT demonstrated what's possible with modern AI, Finn showed us exactly what not to build.

We asked ourselves: "What if banking AI didn't suck?"

What It Does

Instead of another dumb chatbot that gives canned responses, we created an AI that can actually do things:

400+ bunq API endpoints accessible through natural conversation
Real banking operations - check balances, make payments, review transactions
Actually intelligent - understands context, handles complex requests
Secure by design - API keys protected, sensitive operations handled carefully

While Finn struggles with basic customer service, Bunq Atlas can parse your transaction history, execute banking operations through conversation, explain financial concepts in plain language, and handle multi-step workflows without losing context.

How We Built It

Frontend: Next.js 14 + React + TypeScript + Tailwind CSS
AI: Claude 3.5 Sonnet via Anthropic SDK with custom system prompts
Banking: bunq API via MCP (Model Context Protocol) server
UI: Radix UI components with bunq-inspired styling
State Management: React hooks for conversation and banking context

Challenges We Ran Into

MCP Integration: Getting the Model Context Protocol to properly expose 400+ bunq endpoints to Claude required careful tool schema design
Security: Ensuring API keys stay protected while allowing the AI to execute real banking operations
Context Management: Keeping track of multi-step banking workflows (e.g., "transfer €50 to Mom") across conversation turns
UI/UX: Making complex banking data readable and actionable in a chat interface

Accomplishments That We're Proud Of

Built a working AI banking assistant in 24 hours that actually executes real transactions
Successfully integrated 400+ bunq API endpoints through MCP
Created an interface that makes banking feel conversational instead of bureaucratic
Proved that banking AI doesn't have to be as useless as Finn

What We Learned

MCP is incredibly powerful for giving AI agents real-world capabilities
Banking APIs are complex, but good AI design can hide that complexity from users
Context management is everything - an AI assistant is only as good as its memory
There's a massive gap between what's technically possible (Grok, ChatGPT) and what's actually deployed (Finn) in fintech