Ten small apps and one question
What interface exists today only because the user has to configure the software themselves?
The apps I like using tend to have one good screen. A calendar I can read at a glance. A countdown I can see from across the room. A departure board.
Most of the effort in those apps goes somewhere else: into the screens in front of the good one. Set-up wizards, pattern builders, settings, filters, report pages. They exist because the software has to be told what to do, and until recently the only way to tell it was to fill in a form.
Project Titan is an experiment in taking those screens away.
Where it started
Titan didn’t start as a plan. It started with a few small apps I built for myself and the people around me.
- Traintrackr puts live UK train times inside an AI chat. You ask Claude about the next train and the answer, sometimes a departure board, arrives in the conversation.
- Tally runs interval workouts hands-free on a phone. Building a session and asking questions of the log happen in Claude. The phone keeps the clock.
- Gondola lets a group of friends say which resort they’ll be in, so they can see who’s skiing where.
In the first two, the AI ended up doing the jobs that would normally need their own screens, and the app was left with the one screen worth having. The apps got smaller and, I think, better. Titan is my attempt to find out whether that holds up when you do it on purpose, again and again.
Why now
The thing that makes this practical is MCP, the Model Context Protocol. It lets the AI you already use read and change an app’s data on your behalf. Instead of a workout builder, you can say “give me a 20-minute core session with what I’ve got”. Instead of a reports page, you can ask what you trained most this month.
Most of what gets called AI in software goes the other way: a chat box added to an app that already had too many screens. Titan starts from the opposite thought.
AI doesn’t necessarily mean adding more intelligence to software. It might allow us to remove interface.
The rules
Every Titan app follows the same few rules.
- One job. It keeps only the screen that beats a chat: a calendar, a countdown, a map.
- No AI inside. No assistant, no prompt box. The thinking happens in the AI you already use, connected through MCP.
- Open from day one. Anyone can sign up when it launches. It’s free, and it’s an experiment with no guarantees.
- One story. Each app gets a short post arguing the one rule that makes it what it is.
Ten apps, one at a time
There are ten apps in the run, each with its own name and its own address on ultraniche.ai. They’re revealed one at a time on the homepage, like an advent calendar: launched apps are open, the next one shows its name, and the rest are just numbers. I won’t say what an app does until it launches.
The shared kit
Ten apps is only sensible if each one costs little more than its own idea. So before the first one there’s Project Zero: a template every Titan app starts from, and a command that sets each one up and puts it live. Sign-in by emailed link or passkey, an app you can install on your phone that works offline, and an MCP connection are all in the template. Anything every app needs lives there, and what one app learns goes back in for the next. A lot of it came from Traintrackr and Tally in the first place.
The apps themselves are built with Claude Code.
What I want to find out
- Whether taking the configuration screens away makes an app better, not just smaller.
- Where the line falls: which screens survive, and which turn into a sentence.
- What happens when the AI gets it wrong, and what an app needs so you can check its work before anything is saved.
- Whether a small, open experiment can still be built well enough to trust with your data.
Each app’s story will say what held and what didn’t.