← Back to docs
Recipe

Recipe: Personal finance assistant

Build a local-first budgeting agent that categorizes transactions, tracks spending goals, and surfaces insights — all running on-device with zero cloud dependency.

Overview

This recipe wires Meridian into a personal finance workflow. The agent ingests CSV exports from your bank, classifies each transaction using a local embedding model, and maintains a running budget against categories you define. All data stays on your machine.

Prerequisites

  • Meridian CLI v2.1+ installed and authenticated
  • At least one bank CSV export (Chase, Wells Fargo, or generic format)
  • Python 3.11+ for the local embedding runtime

Step 1 — Scaffold the agent

meridian init finance-agent --template csv-classifier

This creates a project with a pre-built CSV parser, a category schema, and a local SQLite store for transaction history.

Step 2 — Define categories

Edit the generated categories.yaml to match your spending buckets. The agent uses fuzzy matching against merchant names and transaction descriptions.

Step 3 — Ingest transactions

meridian run finance-agent --import ./exports/march-2026.csv

The agent classifies each row, flags uncategorized items for review, and updates your monthly budget tracker.

Step 4 — Query insights

meridian ask finance-agent "How much did I spend on dining in March?"

Natural-language queries run against the local transaction index. No data leaves your device.

Next steps

Add recurring-transfer detection, set up weekly summary digests, or chain this agent with the investment tracker recipe for a full wealth dashboard.