Why I built Povver

Arithmetic first, then the model

Povver started as a training app. The coach came later. The boring work had to be right first: hard sets per muscle, an estimated 1RM per lift, your volume against the published ranges, and whether you have earned a heavier weight. That is arithmetic, and it should give the same answer every time.

A language model is bad at arithmetic and good at the part after it. It reads the shape of a training block, weighs a stalled lift against a week you cut short and explains a decision in a sentence you can argue with.

So the app does the counting, and the model reads what the app counted. Both are open over MCP, the standard way to plug data into Claude and other assistants, so any assistant can read the summary, dig into single sets when it needs to, and write the change back into your routine.

How I got here

In 2024 I tried to get coached by a chat window. It nearly worked.

Every week I did the same thing: I exported the watch data and the training log, added what the scale said, pasted it all into a chat, explained the plan and asked what to do next. The first answers were good.

Then it drifted. By about the fifth message it would contradict something I had pasted the week before, or describe a lift I had never done. The reasoning was fine. The old material had just fallen out of the window.

A model only knows what is in front of it right now, and that space is limited. A year of lifting is thousands of sets, before you add sleep, HRV and daily weight. It does not fit. So you paste a slice, the model fills the gaps with a guess, and every new chat starts from nothing. You end up as the memory, and the memory was the part I wanted to hand over.

My first fix was to paste more. That made it worse. The window filled with raw rows, the oldest weeks dropped off first, and the model was doing the adding up.

Povver reads training only. I still read the watch and the scale next to it in Claude, over MCP (the example is in the docs).