My company runs on three bank accounts across two countries and four currencies. For the first year, all of it lived in a Google Sheet. One person updated it manually. Sometimes they forgot. Sometimes the numbers didn’t add up. Sometimes I’d ask “how much did we spend on marketing last month?” and the answer would take twenty minutes.
I replaced the spreadsheet with a full finance system — dashboard, Telegram bot, automated reports, AI analyst — in a weekend. I don’t know how to code.
The spreadsheet problem
Every small company starts with a spreadsheet. It works until it doesn’t.
Here’s when it stops working: when you have multiple accounts in different currencies. When two people need to enter data. When you want to ask a question and get an answer in seconds, not after someone opens a laptop, finds the file, and scrolls through tabs. When you realize your “system” is one person’s memory plus a Google Sheet that nobody trusts.
I didn’t even bother with Quickbooks or Xero. They’re built for standard businesses with standard invoices and standard payroll. My business isn’t standard. Multiple currencies, multiple countries, income flows that don’t map to anything off-the-shelf. I knew from the start that no SaaS tool would speak my language — literally or figuratively.
So I built my own.
What I built
A finance dashboard that does five things:
One — shows real-time balances across all accounts, converted to USD. Two — records every transaction with category, currency, and exchange rate. Three — compares monthly budget plan versus actual spending. Four — tracks team salaries and active subscriptions. Five — lets me ask questions in plain language in a Telegram group and get instant answers.
The whole thing runs on four components: a React web app for the visual dashboard, a PostgreSQL database for storage, a Node.js API server, and a Telegram bot powered by Claude Haiku that understands natural language.
Total monthly cost: about $7. That’s the server. The database is free tier. The bot costs fractions of a cent per message.
How (for non-coders)
I used three AI tools and zero code knowledge.
Step 1: The database. Supabase — a hosted PostgreSQL database with a visual interface. I described my tables in plain English to Claude, got the SQL, pasted it into Supabase’s SQL editor. Accounts, transactions, budget items, team members, subscriptions. Took an hour.
Step 2: The dashboard. I started with Variant — an AI design tool. Typed one prompt, got a full visual layout for a music label finance report. Then I wrote the actual code with Claude and brought it into Lovable — a platform where you can deploy, host, and iterate on the app in a chat-based interface. Need to tweak a component? Just describe the change. Want to plug in an AI feature? It’s right there. Connected it to my Supabase database, and the whole visual layer was done in an afternoon.
Step 3: The brain. This was the interesting part. I needed an API server that could talk to the database and run a Telegram bot — but I also wanted AI assistants to plug into the same data. So Claude wrote the entire server as an MCP server (Model Context Protocol — an open standard that lets AI tools call your functions directly). Every finance operation — check balances, add transactions, pull reports — is both a REST endpoint and an MCP tool. I described what I wanted in conversation. Claude wrote the code. I deployed it to Railway by connecting my GitHub repo. Auto-deploys on every push.
Step 4: The bot. A Telegram bot that sits in our company group chat. Anyone on the team can type “how much did we spend this month?” and get an answer in two seconds. Or “add expense $50 for marketing” and it records the transaction, confirms, done. It uses Claude Haiku under the hood — fast, cheap, understands Russian and English.
The whole system was built in roughly a weekend of focused work. Not a weekend of coding — a weekend of conversations with AI, describing what I want, and deploying the results.
What I learned
Vibe-coding is real, but not magic. You still need to understand what you’re building. I didn’t write the code, but I made every design decision — by answering Claude’s clarifying questions, asking it to check the system for consistency, and requesting decision matrices with pros and cons when choices got complex. The AI writes the code and proposes the architecture. You validate the logic and make the calls.
Design the system before you build it. Before writing any code, I had Claude build a detailed spreadsheet — the exact tabs, categories, account structure, and formulas my business actually needs. I described what I wanted, Claude proposed the structure, I tweaked it in conversation until it was right. That Excel file became the blueprint for everything. When I moved to code, I wasn’t guessing — I was translating a system I already understood. The first code version still had no security and no input validation, but the data model was right from day one. I hardened everything later — authentication, access controls, security audit — also with AI.
The Telegram bot is the killer feature. The dashboard is nice, but I open it once a week. The bot I use every day. Quick entries, quick questions, no context switching. If you build one thing, build the bot.
You will hit walls. Currency conversion logic broke twice. The exchange rate calculation had a subtle bug that took hours to find. A transaction type I added later didn’t play nice with the monthly summary view. AI helps, but debugging is still debugging. The difference is you debug in conversation, not in a code editor.
$7/month beats $50/month. Every SaaS finance tool charges $30–100/month and still doesn’t do what I need. My custom system costs less than a coffee and does exactly what I want — because I designed it for my specific business. Supabase free tier, Railway ~$5, Telegram free, Claude API for the bot $1–2. Lovable is free for building; hosting can be moved to Vercel or Netlify for $0.
The stack (if you want to replicate)
You don’t need to know what these are to use them. You just need to know they exist.
Supabase — free database with a nice UI. Your data lives here. Lovable — AI builds you a web app from a text description. Your dashboard lives here. Railway — hosts your server for ~$5/month. Your API and bot live here. Claude — writes all your code, powers your bot, and becomes your finance analyst through MCP. Telegram Bot API — free. Your quick-entry interface lives here.
I put together a prompt pack with clean, distilled prompts for each stage — based on what I learned building the real thing through dozens of iterative conversations with Claude. My actual process wasn’t messy — I had full project documentation, system prompts, and a structured SDLC from the start. But this prompt pack compresses all of that into something you can pick up and run with immediately.
Download the Finance Stack Prompt Pack →
One thing you can do today
Open Claude. Describe your current finance tracking setup. Ask it to design a database schema for your specific business — your accounts, your currencies, your expense categories. Don’t worry about building anything yet. Just see the schema. That’s your blueprint.
The gap between “I track money in a spreadsheet” and “I have a real finance system” is smaller than you think. It’s a weekend.
Your move.