FocusNow AI Club, meet-up #7 – 28 October 2026.
The seventh session was one demonstration and a long discussion. One of us showed a set of small applications, all built with Claude, that now run his own infrastructure. The room then debated what that means for software.
From cars to clothes
The analogy that framed the session: software used to be like cars. You bought what a manufacturer built, lived with its standards, and changing it took time and money.
AI-assisted development is making software more like clothes. You can own many, cheap or expensive, change them often, and throw one away when it no longer fits.
What one person built
Each tool, opened from a single launchpad, does one job:
- A backup dashboard that shows every server’s daily backups and restores a whole server, or a single app, from one screen.
- A monitoring system with an IT-admin agent that checks devices and services, tries to fix what it can and sends a phone notification about the rest.
- A self-hosted password vault.
- A local git server holding all documentation. The agents read it before starting any task, which noticeably cut the tokens they spend investigating.
- An invoicing app that replaced a paid subscription.
The only running costs are servers and AI subscriptions.
Two models, one builder
He sends each design to a second AI model and puts its view back to Claude, which either defends its approach or adopts the better one. Then Claude writes all the code. The value is in the second opinion.
Routine work stays off the largest models. Small, fast models run the monitoring agents well enough, and cost less.
Managing ideas
The last tool was an idea pipeline. It is a bit similar to SPM Demand Management, the key difference – AI acting as devil’s advocate. Each idea is challenged against written context: strategy, decision log, 12-month goals and constraints. It is then scored, ranked and moved between stages.
On MCP, the room agreed: in our own CRM, data entry now happens by asking Claude, while dashboards are still viewed in the app. Claude is becoming the front door to the application.
Key takeaways
- Small, disposable tools are now practical. One person built and runs a set of them for the price of hosting and AI subscriptions.
- Give agents documentation first. Context read before starting saves tokens.
- Use a second model as a reviewer. Disagreement between models is useful input before anything gets built.
- Ideas still decide value. If anyone can build software, what sells is the idea and the quality, just as people still buy good clothes.
Join the next one
The FocusNow AI Club meets monthly, with working demonstrations rather than slideware. If you are deciding which small tools to build for yourself, get in touch.
Sources: Vaultwarden wiki · Gitea · ServiceNow Demand Management workflow