The era of personal software
📝 摘要
作者以自己构建的健康数据追踪应用为例，讨论 LLM 时代「个人软件（personal software）」的兴起：当构建成本足够低，软件将逐渐商品化，越来越多人会为自身需求量身定制工具。他完整记录了实现链路——Apple Health 数据经 Health Auto Export 每小时导出 JSON，通过 ingest 端点写入 $5 的 PlanetScale Postgres；数据库 schema 与历史数据回填由 Agent 通过 MCP 自动完成，首版约 30 分钟上线。AI 部分最初用 AI SDK + Claude Haiku 5.5 做餐食热量估算，但因额外花费、交互方式和准确率问题放弃，改为开放 /api/meals 端点，让 Grok Bot、Claude、Codex 等已有订阅的 Agent 直接写入，并兼任健康教练（补剂提醒、周度洞察）。设计上采用 markdown 风格，包含 Today 概览、GitHub 式活动格、四类图表和明细表格。作者还建议让 Agent 基于历史对话找出可构建的个人工具，并列举了 oklch.fyi、Loading.dev、design.interfere.com、Meniny 等例子。
核心概念： 个人软件并非新概念，真正变化的是构建门槛——模型让「只服务自己、无需向他人解释」的定制软件变得经济可行。
关键技术栈 / 关键要点：
- 数据采集 — Health Auto Export 每小时把多设备健康/运动数据以 JSON 推送到 ingest 端点。
- 数据库与 MCP — PlanetScale Postgres（$5）通过 MCP server 让 Agent 自动建表、写 schema、回填历史数据。
- Agent 集成 — 放弃自建 AI SDK Agent，改为开放写接口，复用 Grok Bot / Claude / Codex 订阅做热量估算与健康教练。
- 设计 — markdown 风格界面 + Today 概览 + GitHub 式活动图 + Steps / Active Energy / Eaten / Sleep 图表。
- 找想法 — 让 Agent 检索过往对话，找出手动重复、反复询问的事项作为构建候选。
关键结论 / 实践建议： 不必替换所有现成工具，只针对高频且有明确个人偏好的场景自建；个人软件的价值在于允许「奇怪」的取舍，例如不满足对比度检查但自己喜欢的配色。
Something that keeps coming up over and over again, pretty much since LLMs started getting good at building things, is the idea of personal software.
The idea is that once models get good enough, the effort to create an app that suits your exact needs will be so low that it will eventually replace most traditional software and software will become a commodity.
com·mod·i·ty /kəˈmɒdɪti/
- 1an article of trade or commerce, especially a product.
- 2something of use, advantage, or value.
- 3 Econ. a good or service whose wide availability typically leads to smaller profit margins and diminishes the importance of factors (such as brand name) other than price.
I don’t believe we’re there yet but there are more and more people building their own personal tools that they use both at work and in their personal lives.
I recently built an app I’d call personal software, one that I use every single day and it even made its way to my homescreen. I thought it would be interesting to document the process from start to finish and write about it.
My homescreen with the app I built
The idea
I’ve been spending a lot of time in NYC recently and I’m walking around way more than I used to. I also started tracking my exercise and I wanted to create a place where information from all of my devices would be stored and displayed.
|
|
|
|
|
___|___ |
| | | ___
| | | |_|_|
| | _|_ |_|_|
| | |=| |_|_|
|\ /| _|=|_ | |
/:\ /:\ _|:::|_ | |_|_|
|:\ /:| _|:::::|_ | |_|_|
|:\ /:| |:|:|:|:| (^) |_|_|
|:\ /:| |:|:|:|:| ((^)) | |
____ /::\ /::\ |:|:|:|:| (((^))) |_|_|
|::::| |:::\ /:::| |:|:|:|:| <|:|:|:|> |_|_|
|::::| |:::\ /:::| |:|:|:|:| |:|:|:| ___ |_|_|
|::::| |:::\ /:::| ____ |:|:|:|:| |:|:|:| |:::| | | ____
____|::::| |:::\ /:::| |::::| |:|:|:|:| ___ |:|:|:| |:::| |_|_| |::::|
|::::|::::| /::::\ /::::\ |::::| _|:|:|:|:|_ |:::| __|:|:|:|__ |:::| |_|_| |::::|
|::::|::::| |::::\ /::::| |::::| ___ |:|:|:|:|:| |:::| |:|:|:|:|:| |:::| |_|_| |::::|
|::::|::::| |::::\ /::::| |::::|___ |:::| _|:|:|:|:|:|_ |:::| |:|:|:|:|:| |:::| | | |::::|
|::::|::::| |:::::V:::::| |::::|:::|____|:::| |:::|:|:|:|:::| |:::|_|:|:|:|:|:|_|:::| |_|_| |::::|
|::::|::::| |:::::::::::| |::::|:::|::::|:::| |:::|:|:|:|:::| |:::||:|:|:|:|:|:||:::| |_|_| |::::|
|::::|::::| |:::::::::::| |::::|:::|::::|:::| |:::|:|:|:|:::| |:::||:|:|:|:|:|:||:::| |_|_| |::::|
|____|____|_|___________|_|____|___|____|___|_|_____________|_|___||___________||___|_|_|_|_|____|
It’s not exactly a ground-breaking idea and there are definitely apps I could’ve used, including Apple Health, that would do an okay job.
But none did exactly what I was looking for. I wanted a web app with a very specific design. I’ve also wanted to try PlanetScale for a while now and this felt like the perfect opportunity to do so.
The structure
The app is quite simple since I already had all of the data on my devices. The remaining problem to solve was getting it from my devices into the app and figuring out how to display it.
iOS automatically syncs the data from your devices to the native Health app and allows you to export it, which is great. I was only interested in a small subset of that data and I wanted the export to be automatic, ideally at regular intervals.
Apple Watch
Garmin watch
iPhone
Whoop
Health app
I found an app called Health Auto Export that lets you set up an automation and specify exactly what data you want to export. I set it up to export the workout and health data as JSON every hour.
I could’ve probably built this part too, but since it wasn’t the focus, it felt more reasonable to just let another app that specializes in this handle it.
The database
As I mentioned earlier, I wanted to use PlanetScale for the database, specifically the $5 Postgres, since that would be a great fit for an app like this. I onboarded using the MCP server and let an agent set everything up. It was incredibly easy.
This is something I’m experiencing more and more when onboarding to new tools. I don’t really interact with the UI at all, the agent just uses an MCP to set everything up on its own.
PlanetScale dashboard for the health database
It set up the schema and a script to backfill all of my historical data. The automation sends data to an ingest endpoint every hour.
The database has 5 tables: workouts, sleep and meals, samples for daily totals like steps and energy and ai_usage, to track the AI spend.
- workouts
- sleep
- meals
- samples
- ai_usage
AI Agent
After deploying the first version in about 30 minutes, I realized that in order to make this actually useful, I’d need to add meal tracking too. The easiest way to do that was to add an agent that can estimate the calories for each meal I add.
| Day | Food | Kcal |
|---|---|---|
| 10-29 thu | Tomato bagel from Apollo | 450 |
| 10-29 thu | Iced latte from Variety | 190 |
| 10-28 wed | Cheeseburger at Hamburger America | 650 |
| 10-27 tue | Margherita at Lucali | 1,200 |
| 10-26 mon | Steak over rice from Inday | 850 |
1–5 of 112
I built the agent with the AI SDK and while it was extremely easy to set up, I quickly realized that it wasn’t what I wanted.
const result = streamText({
model: "anthropic/claude-haiku-5.5",
instructions,
messages: await convertToModelMessages(messages),
stopWhen: isStepCount(15),
tools: {
webSearch: gateway.tools.parallelSearch({ maxResults: 5 }),
logFood: tool({
description,
inputSchema: z.object({ day: mealDay, items: mealItems }),
execute: ({ day, items }) => addMeals(day, items),
}),
},
});
Paying for extra AI spend when I already had multiple subscriptions felt silly and I figured I should just use those instead.
It was also obvious that having to open the app every time I wanted to log a meal wasn’t the experience I wanted. Getting accurate values from a cheap model like Claude Haiku 5.5 turned out to be pretty difficult too.
Health Bot chat logging a meal of bagels with cream cheese, turkey ham and mozzarella
Instead, I added an endpoint so agents like Grok Bot, Claude or Codex can write to the table. They’re much better at estimating values and need no extra spend.
{
"items": [
{ "name": "2 bagels with cream cheese", "kcal": 620 },
{ "name": "2 turkey ham slices", "kcal": 60 },
{ "name": "2 mozzarella slices", "kcal": 160 }
]
}
I created a bot called Health in Grok Bot and gave it a token so it can write to the database. I send it the meals I eat, it researches the values online and logs them through the endpoint.
It has the entire chat history too, which is useful when things get repeated or it needs context.
Auto Export
Ingest
Health Bot
/api/meals
PlanetScale
Health App
It also reminds me to take supplements, sends me insights from the past week and keeps an eye on long-term trends. It’s basically my health coach.
Health bot routines for a daily supplement reminder and weekly and monthly health insights
Design
When it comes to the design, I had a clear idea of what I wanted. I’ve been enjoying the markdown-esque aesthetic recently, so I wanted to do something in that realm. It also works well for charts and tables, which this app is full of.
Health log page showing today's stats, a weekly activity grid and a steps chart
At the top of the app the Today section gives me an overview of my daily stats, like how I’m doing with my calorie tracking and how many steps I’ve taken.
Below, there’s a GitHub-like activity section that shows me my exercises and Active Energy, Steps, Eaten and Sleep charts.
00:0000:03
At the bottom of the page there are a couple of tables that show me the past 10 entries for each category with the option to see more.
Sleep table listing hours slept, bedtime and wake time for each day
The beauty of building personal software is that you can do anything you want. A health data gathering web app with markdown-esque design is not something I’d imagine would be high up on people’s desired apps list.
GitHub repository page for the private health project
Finding ideas
Previously, only huge companies had the resources to build internal tooling for their specific needs, now it feels like anyone can do it.
Claude Code prompt asking what software I could build for myself that I'd use every day
It can be difficult to come up with what you should build though. You shouldn’t build just for the sake of building. If you’re unsure, I’ve found asking your agent to be a good starting point.
Based on everything you know about me, suggest a few personal apps or tools I could build that I’d actually find useful. Look for recurring tasks, things I track by hand, questions I keep asking or tools that don’t quite work the way I want. For each idea, explain what it would do and what from our conversations made you suggest it.
There’s an incredible amount of insight in your past conversations, especially things you’ve been doing manually or asking about over and over again.
Recents
Daily running plan
Calories in a tomato bagel
Is my resting heart rate normal?
Weekly training load summary
Calories in an iced latte
Garmin vs Apple Watch sleep data
Protein in two turkey ham slices
Why is my HRV lower today?
Calories in steak over rice
How many steps did I walk this week?
For example, a lot of the skills I’ve built started off like this. I asked my agent what kind of things I was doing manually over and over again and then built skills to do them for me. I originally didn’t plan on publishing them and only did so later.
More examples
A few other projects I’ve built also started this way. I built oklch.fyi because I needed a place where I could play around and test OKLCH colors. I only made it public later after seeing the demand on Hacker News.
Loading.dev was the same. We built it, because we were tired of having to define the same spinners in multiple projects over and over again. We only made it public after we’d been using it in our projects for a while.
Same for design.interfere.com. We built it as an internal tool that helps us document and work with our design system and only happened to make certain parts of it public later.
Meniny is an iOS app that I haven’t shared anywhere before, largely because it’s only useful to a handful of people. It allows you to track namedays and birthdays easily in a single app. I mostly built this for myself, my friends and my family.
There are a couple other tools and apps that I built recently that have made my workflow much better, some of which have actually helped with this article and I’ll be releasing those soon.
Conclusion
People have been building tools that solve their own problems for a very long time. When solving a problem you encounter, you have a deep understanding of the struggle and are passionate about the solution.
Before models got this good, my Health app would’ve taken me much longer to build and I probably wouldn’t have done it.
It wouldn’t have been worth the time and I would’ve used an alternative built by someone else that didn’t do exactly what I wanted.
When building for yourself, the way you build changes. You no longer need to make decisions that work for everyone and you don’t need the app to make sense to anyone else but you.
If I really like a color combination and it doesn’t exactly pass contrast checks? Oh well, I’ll just use it anyway. There’s no real risk to making that choice other than affecting my own experience.
It doesn’t mean replacing all your tools with custom ones, at least for now. I wouldn’t do a better job than the person who built Health Auto Export because I haven’t spent enough time thinking about the problem, nor do I want to.
Personal software is not a new idea. What changed is how accessible it has become. You can easily build something tailored to your specific needs, exactly the way you want it.
Just you and your experience, even if it’s a little weird.
More
If you enjoy articles like this and want to learn more, take a look at Interfaces, my design engineering magazine.
It’s where I share everything I know about building great interfaces. From animation and typography to performance and accessibility.
InterfacesThe Design Engineering
Magazine font-libre-baskerville