I built an MCP server that lets Claude manage my personal finances

The problem with every finance app

I’ve tried them all: Mint, YNAB, Monarch, Copilot, Rocket Money. They all share the same fatal flaw: they require you to maintain them.

Transactions need categorizing. Rules need tweaking. Budgets need adjusting. Miss a week and you’re staring at a wall of “Uncategorized” transactions that makes you want to close the app forever. This isn’t just me — it’s the #1 reason people abandon personal finance tools.

The dirty secret of personal finance apps is that they create more work than they save. You traded a spreadsheet for a prettier spreadsheet that sends you guilt-trip notifications.

What if you never had to open a finance app again?

I already talk to Claude and ChatGPT every day. What if they could just… talk to my bank data? No new app to open. No transactions to manually review. Just ask questions and get answers.

“How much did I spend on dining this month?” “What are my recurring subscriptions?” “Will I have enough to cover rent next week?” These are the questions I actually care about. I don’t want to stare at charts — I want answers.

That’s why I built Cashflow: a personal financial ledger that exists entirely as MCP tools for AI agents.

What’s MCP?

Model Context Protocol(MCP) is a standard that lets AI assistants call external tools. Think of it like USB for AI — a universal connector between your AI assistant and the outside world.

Claude, ChatGPT, and a growing list of AI clients all support MCP. When you connect an MCP server, your AI assistant gains new capabilities. Connect a GitHub MCP server and it can manage your repos. Connect a Slack MCP server and it can send messages. Connect Cashflow and it can manage your finances.

The key insight: your bank data becomes just another tool the AI can use. No separate interface needed. The AI assistant you already use every day becomes your personal accountant.

How Cashflow works

The setup takes about a minute:

• Connect your bank via Plaid (12,000+ supported institutions). Transactions sync automatically every hour.
• Add the MCP server URL to Claude or ChatGPT: https://cashflow.tech/api/mcp
• Start asking questions.That’s it. No app to configure, no budgets to set up, no categories to define.

Behind the scenes, Cashflow exposes three tools to your AI assistant:

• query— Read your financial data. Spending summaries, transaction search, recurring bill detection, cash flow forecasting, anomaly detection, and more. The tool auto-detects which mode you need from the parameters.
• annotate— Classify and organize transactions. Set categories, assign parties, tag transactions, mark transfers. Everything is a non-destructive overlay on your raw bank data.
• admin— Manage reference data. Create categories, set up rules for automatic classification, merge duplicate merchants, rename accounts.

What you can actually ask

Here are real examples from my daily use:

“How did I do this month?”— Returns income, expenses, net savings, top spending categories, and comparison to last month. Includes savings rate and housing percentage if you want the ratios.

{
  "period": "this_month",
  "compare": "prior_period",
  "include": [
    "ratios"
  ]
}

“What are my recurring bills?”— Detects subscriptions and recurring charges automatically. Shows frequency, amount, and a forecast of upcoming payments.

{
  "recurring": true,
  "forecast_days": 60
}

“Any unusual charges this month?”— Flags spending spikes, unusually large transactions, price changes on subscriptions, and new merchants with large charges.

{
  "period": "this_month",
  "by": [
    "group"
  ],
  "type": "expense",
  "include": [
    "anomalies"
  ]
}

“Will I make it to the end of the month?”— Projects your balance forward based on recurring income and expenses, showing the lowest point and when it occurs.

{
  "forecast": true,
  "forecast_days": 30
}

The architecture (for the nerds)

Cashflow is a Next.js app deployed on Vercel, with a SQLite database on Turso managed through Drizzle ORM.

A few design decisions that matter:

• Immutable raw data + mutable overlay. Raw transactions from your bank are never modified. All classifications (categories, parties, tags) live in a separate enrichments table. You can always re-derive from the original data.
• Specificity-based rules engine.Transaction rules match on description patterns, and longer (more specific) patterns win. No manual priority ordering — the most specific rule always applies.
• Token-efficient responses. Every MCP response uses terse field names (amt, cat, grp, desc) and omits nulls. This keeps AI context windows lean so you can ask follow-up questions without hitting limits.
• Controller, not CFO. Cashflow focuses exclusively on cash management: income, expenses, bills, transfers. No investment tracking, no net worth, no retirement projections. This focus means fewer tools, simpler responses, and better answers for the questions that actually matter day-to-day.

Try it

Cashflow is live at cashflow.tech. You can explore the demo data without connecting a bank account to see how it works.

To connect it to your AI assistant:

• Claude.ai or ChatGPT:Go to Settings → Integrations → Add Integration. Paste https://cashflow.tech/api/mcp and authorize with Google.
• Claude Code: Run claude mcp add cashflow https://cashflow.tech/api/mcp

The documentation has more details on all three tools and their parameters.

I built Cashflow because I wanted to stop maintaining a finance app and start having conversations about my money instead. If that resonates, give it a try.