Munny

Personal Open Banking aggregator with an AI financial agent.

Aggregates bank accounts through Enable Banking (PSD2, redirect-based SCA consent), stores history in SQLite, classifies transactions with a hybrid rule + LLM pipeline, and exposes everything as a remote MCP server so an AI assistant can query the data directly. A server-rendered dashboard and a weekly pt-PT insight report sit on top.

Architecture

Enable Banking (PSD2)  ──►  app/sources/enable_banking.py   JWT RS256, consent flow, paginated sync
                            app/sync.py                     per-ASPSP daily request quota
                                    │
                                    ▼
                            SQLite  (accounts, bank_connections, transactions)
                            idempotent via UNIQUE(source, external_id)
                                    │
        ┌───────────────────────────┼───────────────────────────┐
        ▼                           ▼                           ▼
app/categorize.py            app/mcp_server.py            app/dashboard.py
keyword rules (categories.yaml)   Streamable HTTP /mcp    FastAPI + Jinja2 + HTMX
+ batched LLM fallback (Gemini)   bearer auth             balance cards, filterable
manual overrides survive re-sync  read + write tools      ledger, spend-by-category
        │
        ▼
app/report.py  ──►  weekly job (APScheduler): insights → Gemini narration (pt-PT) → Resend email

Components

• app/sources/enable_banking.py — RS256 JWT auth, PSD2 redirect consent, balances and transactions with continuation-key pagination.
• app/categorize.py — keyword rules first (categories.yaml), batched Gemini fallback for the rest, with a short rationale per transaction. Manual corrections are marked category_source = 'manual' and survive a re-sync.
• app/mcp_server.py — MCP server (Streamable HTTP) mounted at /mcp, bearer auth, fails closed on an empty token. Tools: weekly_summary, recurring_payments, upcoming_obligations, save_weekly_report.
• app/dashboard.py — server-rendered pages in the same FastAPI app (Jinja2 + HTMX + Chart.js via CDN, no build step). Cookie session signed with HMAC-SHA256 (stdlib).
• app/insights.py — pure functions: weekly summary vs. prior week and 4-week average, recurring-payment detection, upcoming-obligation reserve funding.
• app/report.py — weekly APScheduler job: assembles insights, has Gemini narrate them in pt-PT, stores the report and emails it via Resend.

Run

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # fill ENABLE_BANKING_APPLICATION_ID, ENABLE_BANKING_PRIVATE_KEY_B64,
                       # MCP_BEARER_TOKEN, GEMINI_API_KEY
uvicorn app.main:app --reload

pytest

Deploy

Dockerfile (python:3.12-slim, uvicorn on $PORT). APScheduler runs inside the app lifespan and syncs every 6h — single replica only. Runs on Railway.

Compliance

PSD2 / SCA consent flow, per-ASPSP request quotas, no credentials in the repo (secrets/ and .env are gitignored).