Building a local-first AI cost management app taught me that shipping software is about much more than writing code.
I recently released v1.5.5 of AI Cost Management, a Windows desktop app that helps users review recorded AI spending across providers and models.
Here are five lessons I learned along the way.
1. Be honest about the data
When pricing information is missing or a cost cannot be verified, showing a confident-looking number can be misleading. I learned to treat unknown values as unknown instead of filling gaps with assumptions.
Trustworthy software should make its limitations clear.
2. Local-first is a deliberate choice
Keeping data on the user's own machine can be valuable for an app that handles spending information. But local-first design also means being clear about what data is stored locally and when the app makes an external connection.
Privacy is not just a slogan. It should be reflected in how the software works.
3. A successful build is not a finished release
Installation instructions, download links, required files, and first-run behavior all affect the user experience.
An installer may require multiple files to be kept together. If the download page does not explain this clearly, users can get stuck before they can even try the app.
I now consider release instructions and download verification part of the product itself.
4. Activity is not the same as cost
Seeing an AI-related process running does not tell you exactly how much it has spent. Process activity, recorded usage, pricing information, and verified costs are different things.
Keeping those distinctions clear helps avoid misleading claims.
5. Real feedback matters more than endless planning
It is tempting to keep adding features before releasing a product. But shipping a defined version creates an opportunity to learn from actual users instead of relying only on assumptions.
My current focus is gathering feedback and finding out which improvements would make the biggest practical difference.
About the project
AI Cost Management is a local-first Windows desktop app for reviewing recorded AI spending across supported providers and models. It is not a universal real-time cost monitor, and costs are shown only when the available data supports them.
Project page: https://jademoon0428-star.github.io/ai-token-cost-monitor/
For other developers: What is one lesson you learned about shipping or maintaining software that you did not expect before real users tried it?
Top comments (1)
I'm especially interested in hearing from developers who maintain small, independent projects. What has helped you get useful feedback from your first users?