The most valuable AI use case isn't a product. It's a personal tool.

There is a version of AI that gets a lot of attention and a version that does not. The one that gets the attention is commercial: AI copilots, AI add-ons, AI agents that companies can license. The one that does not is quieter and, honestly, more genuinely useful.

It is people building their own small tools that connect directly to the browser-based apps they already live in every day. Run them locally on their own machines. Drop a keyboard shortcut on the desktop to access them in seconds. No new subscriptions. No vendor lock-in. No waiting for a product team to build the feature you actually want.

I have been watching this pattern grow for the past year, and what strikes me is how many different kinds of people are doing it. The insight driving them is usually the same: you already have a set of tools that run your actual work and life. Gmail. Google Calendar. Notion. Slack. Todoist. And most people are still interacting with those tools entirely by hand, the same way they always have.

What the more technically inclined figured out is that you can build a bridge between an AI and those tools, run it on your own machine, and create a desktop launcher to trigger it whenever you need it. Then they went ahead and did it.

Simon Willison, the developer behind the open-source Datasette project, built a command-line tool called llm that now has over 12,000 GitHub stars. It lets you send files, clipboard contents, or any text directly to major AI models from your terminal. His personal tools repository has over 77 small apps and scripts, most of them AI-generated, and he adds new ones at a rate of several per week. That pace does not come from a chat interface. It comes from having purpose-built tools wired into an actual workflow.

Marc Bara documented a more elaborate version earlier this year in a detailed Medium post. He built what he calls MarcOS: a personal system running on SQLite that syncs bidirectionally with Gmail, Google Calendar, and Outlook, with a local HTML dashboard that costs zero AI tokens to read. His operating principle is blunt: "Every operation I do more than once a day should cost zero tokens." Claude Code handles the judgment-requiring work. Everything repetitive runs locally.

Alex Honchar connected Claude Code to Gmail, Google Calendar, Slack, Notion, and Clay and built something he describes as a personal AI Chief of Staff. The system generates weekly briefs from upcoming meetings, pulling in relevant context from each of those tools automatically. Where no native integration exists, he built a Chrome Extension to bridge the gap. The whole setup is documented on GitHub and Medium.

Talha Tahir put a dollar figure on his version: $26 a month total. A $20 Claude Pro subscription plus a small VPS. His AI assistant runs 24 hours a day, delivers daily calendar briefings, and uses Discord as the chat interface. The machine runs in Helsinki and he accesses it from anywhere.

Stephen Jayakar's take was simpler but pointed. He had 451 Notion journal entries he wanted auto-tagged. Notion's AI add-on costs $8 a month on top of the base subscription. He built his own tagger using GPT-4o and the Notion API, published the whole project on GitHub, and got the same result for almost nothing in API costs. His framing was direct: why pay for a feature when you can build it in a week and own it outright?

The technology making most of this possible is Anthropic's Model Context Protocol, or MCP. It is an open standard released in late 2024 that lets Claude and other AI systems connect to external services in a structured, consistent way. Since it launched, developers have built community MCP servers for nearly everything. Taylor Wilson's Google Workspace MCP has over 2,600 GitHub stars and gives Claude full access to Gmail, Calendar, Drive, Docs, and Sheets. Nate Spady's standalone Google Calendar MCP has over 1,100 stars and installs in Claude Desktop with a single line in a config file. The tooling has gotten genuinely accessible to anyone willing to spend an afternoon on it.

On the desktop launcher side, Raycast has become the go-to tool on macOS for wiring all of this together. The app, which raised a $30 million Series B in 2024, has built-in AI with Claude, GPT, and others, plus an extension marketplace covering Notion, Todoist, Linear, GitHub, Slack, and dozens more. You can write a custom AI command, bind it to a hotkey, and trigger it on selected text or clipboard content from anywhere on your machine. The community-built raycast-g4f extension alone has over 1,100 GitHub stars. Robert Oberg published an Alfred workflow called Kiki that does something similar for that launcher, with hotkey triggers, text transformation presets, and voice input via Whisper, all running locally on macOS.

For a more complete desktop application, PyGPT is an open-source app running on Windows, Mac, and Linux that connects to Gmail, Calendar, and Slack through plugins while supporting every major AI model including local ones. Leon, another open-source personal assistant with over 17,000 GitHub stars, runs entirely on your own machine with local model support and a skills system you extend yourself.

Paco Cantero built something he calls Mindset: a personal knowledge assistant running on SQLite with twelve specialized agents and a Slack bot as the interface. His description of it captures the appeal of this entire category well: "It is fast, private, and entirely mine." The system never sends his data anywhere he has not explicitly approved.

What all of these projects share is ownership. The builder controls the tool. There is no vendor who can raise the price, deprecate the integration, or change what data gets sent where. The cost is transparent. The behavior is exactly what you configured it to be.

The commercial AI tools have their place. Some of them are genuinely useful. But the people getting the most out of AI right now are not primarily the ones subscribing to the most AI-enhanced SaaS. They are the ones who spent a weekend building something specific, wired it into the tools that already run their day, and dropped a shortcut on their desktop.

That is a different kind of leverage. And it compounds in a way that a monthly subscription never really does.