How To Build Finance Apps With Claude Code

I want to show you something that changed how I think about AI tools for finance.

Most AI tools for finance follow the same pattern. You type a question, you get an answer or some generated code, and then you copy it somewhere and spend the next twenty minutes figuring out why it doesn’t work. The AI gave you instructions. You’re still doing the work.

Claude Code is different. Not in a marketing-language way — structurally different. It doesn’t generate code for you to run. It runs the code itself. It writes it, executes it, reads the output, catches errors, and fixes them. You describe what you want. It builds it. That’s the shift that matters for finance professionals.

What Is Claude Code?

Claude Code is Anthropic’s agentic coding environment. The word “agentic” gets thrown around a lot, so let me explain what it means in practice.

A regular AI chatbot responds to prompts. You ask, it answers. Whatever happens next is on you.

Claude Code operates in a loop. It receives your instruction, writes code to accomplish it, executes that code in your project folder, reads what comes back, and adjusts if something went wrong. The loop runs on its own. If it hits an error — and it will hit errors sometimes — it diagnoses the problem, rewrites the relevant section, and tries again. You don’t have to intervene.

For finance professionals, the implication is significant. The bottleneck has never been access to AI-generated code. It’s been the gap between “here’s a script that should do what you want” and “here’s a working, running tool you can actually use.” Claude Code closes that gap.

It’s also important to understand what it’s not. Claude Code is not a replacement for understanding your data or your process. You still need to know what you want, why you want it, and what a correct output looks like. The tool executes. The judgment is still yours.

What You Need To Run Claude Code

The requirements are simpler than you might expect.

You need the Claude desktop app, which you can download from claude.ai. You need a Pro, Max, Team, or Enterprise subscription — the Code tab isn’t available on the free plan. And you need a folder on your computer where Claude Code can work. That’s it.

No Python installation. No Node.js. No terminal commands. The desktop app includes everything Claude Code needs to run locally. If you’ve been put off by AI coding tools that require a developer setup before you can even start, this is worth revisiting.

Setting Up Claude Code

Once you have the Claude desktop app installed and you’re signed in, click the **Code** tab at the top of the window. You’ll see three tabs: Chat, Cowork, and Code. Code is what we’re using.

Click **Select folder** and choose a project directory. I’d recommend creating a folder specifically for this — call it something like `finance-projects` or `claude-code-work`. This folder is where Claude Code will write files, execute scripts, and save outputs. Think of it as the workspace it has access to.

Once the folder is selected, you’re ready to start. There’s nothing else to configure.

One thing worth understanding upfront: when you send Claude Code a prompt, it will show you exactly what it plans to do before applying any changes. You’ll see a diff view — a before-and-after comparison of what it’s about to write or modify. You approve or reject each change. By default, nothing gets written to your files without your say-so. That’s a meaningful safety net when you’re working with real financial data.

Build 1: Cleaning and Visualizing a Finance Dataset

Let’s start with something immediately useful — a task that most finance professionals deal with constantly.

You have a CSV file exported from two different systems. Department names are inconsistent (“Marketing” in one row, “marketing” in the next, “Mktg” in a third). Dates are formatted differently across rows. There are blank rows from the export. Some actuals values are missing.

Before you can do anything analytical with this file, you have to clean it. And cleaning it manually — standardizing names, fixing dates, handling blanks — is the kind of work that takes 20–30 minutes every time and produces no insight on its own. It’s pure overhead.

Here’s what the prompt to Claude Code looks like:

> "I have a CSV called budget_actuals.csv in this folder. It has inconsistent department names, mixed date formats, blank rows, and some missing actuals values. Please: (1) clean the dataset by standardizing department names, fixing date formats to MMM-YYYY, removing blank rows, and flagging missing actuals rather than deleting them, (2) save the cleaned version as budget_actuals_clean.csv, and (3) create a bar chart comparing Budget vs Actuals by Department for the most recent month and save it as variance_chart.png.

Claude Code reads the file, writes a cleaning script using pandas (a Python data library), executes it, and if the first pass hits any errors — date parsing issues are common — it catches them, rewrites that section, and runs it again. The final output is a clean CSV and a chart, saved to your folder.

From prompt to output, this takes under two minutes. The manual version of the same task is a half hour of work before you’ve analyzed anything.

A few things make this prompt effective. Being specific about what “clean” means rather than leaving it vague. Specifying the output format you want. Naming the output files explicitly. The more context you give Claude Code upfront, the fewer iterations it needs.

Build 2: A Cash Flow Tracker App

This is where Claude Code moves from “useful script” to “actual tool.”

The goal: build a Python app that lets you enter monthly inflows and outflows by category, calculates net cash flow, tracks a running cumulative balance, displays a chart distinguishing forecast months from actuals, and flags any month where the balance goes negative.

This is not a one-off script. It’s something a finance team could run repeatedly, share internally, and build on.

The prompt:

> "Build me a Python app for tracking monthly cash flow. It should: (1) let me input monthly inflows and outflows by category, (2) calculate net cash flow for each month, (3) show a running cumulative cash balance, (4) display a line chart with net cash flow by month and the running balance as a second line, and (5) label any month marked as a forecast differently from actuals — use a dashed line for forecasts vs a solid line for actuals. Save the app as cashflow_tracker.py and make it runnable from the desktop."

Claude Code works through this in layers — data structure, then calculation logic, then the chart. It takes a few minutes. Once it has a first version ready, you can ask it to add features mid-build the same way you’d describe them to a developer: in plain English.

Adding a feature looks like this: “Before we run it, add one more thing — if the running balance drops below zero in any month, highlight that bar in red on the chart.”

That’s it. Claude Code reads its own code, finds where the chart is built, adds the conditional formatting, and updates the file.

When you run the finished app with sample data, here’s what you see: solid lines for actual months, dashed lines for forecast months, the cumulative balance tracking alongside net cash flow, and February flagged in red because that’s where the model dipped below zero.

The thing I want you to notice: cashflow_tracker.py is now sitting in your folder. It’s a real file. Someone else on your team can run it. You can modify it — again, just by asking Claude Code in plain English. You can add new categories, change the chart style, build in export functionality. You did not write a single line of code. You built something that functions like software.

What Claude Code Can’t Do

Worth being honest about the limits.

Claude Code writes and runs code. It doesn’t understand your business context unless you tell it. If you ask it to “clean this dataset,” it will apply reasonable defaults, but it won’t know that your finance team considers a missing actual a critical error versus a minor one. You need to specify those judgments in your prompts.

It also doesn’t have access to your ERP, your data warehouse, or any external systems unless you give it a file or a connection. It works with what’s in the folder.

On data security: Claude Code in local mode runs entirely on your machine. Files in your project folder don’t leave your computer (unless you send them in a prompt). For highly sensitive financial data, this is the setup to use. Remote sessions run on Anthropic’s cloud infrastructure, which introduces different considerations — worth reviewing their data handling policies before using that mode with anything confidential.

The learning curve is real, even if it’s smaller than you’d expect. Writing effective prompts for Claude Code is a skill. The first few attempts will be less efficient than the ones after you’ve calibrated what level of specificity you need. The good news: you get better at it fast, because you can see immediately what works and what doesn’t.

How to Get Started This Week

Two specific things to try, in order.

First: the dataset clean. Find a messy export you already have — anything with inconsistent formatting — and prompt Claude Code to clean it. Don’t start from scratch; use a file you’ve actually been frustrated with. The contrast between before and after is where it clicks.

Second: the cash flow tracker, or whatever the equivalent is for your workflow. Pick something you’ve thought “I wish I had a tool for this” and describe it to Claude Code. Start simpler than you think you need to. A basic version that works is better than an ambitious version that takes five iterations to get right.

The framework for writing good prompts: state the inputs, describe the transformation or calculation, specify the output format and file name. That covers most finance automation tasks.

If you want to go deeper, Finance Automation Lab is where I work through builds like this with a community of finance professionals — including the prompt library and templates from the demos in this article.