Tech-Blog
Acopalypse
Mipsy and Roller are âtheâ two apps Iâve had Claude Opus 5.5 implement for me to specification in an interactive session using their subscription product (October 2026). From my earlier testing over the years, using OpenAI and Gemini models, Iâve noticed a categorical jump in capabilities, so markedly that I donât feel I have to expand on it - if you are a developer, I think you will know.
I donât intend to write more apps in this way, it is not healthy and has a decoupling, dulling effect, mixed with a slightly addictive component. The workflow disconnects you from implementation, experimentation and learning. I chose these two apps expressly so that they will serve me as tools for leveraging what I personally think AI has to offer me as an individual and knowledge worker, and call it quits on online SOTA models.
- Mipsy - The text machine
If you know Acme (Rob Pikeâs editor for Plan 9), Mipsy is a teeny-weeny bit like it. It is basically an app for producing text. You can pull text around into boxes, generate it using scripts, and do LLM inference with it to make and consume more text.
It can work with ZIM files, so you can search a local copy of Wikipedia or Stack Overflow almost instantaneously for keywords and phrases, and drag the result (including autogenerated references) into LLM context boxes to refine them. Using a tree-view, you can organise your text in your local file system as text-files into which you can drag-and-drop, or in the appâs SQLite database. The text nodes in the SQLite store have a token-compaction feature: when you drag a parent node, the child nodes are concatenated. But you can tick a box on the parent-node, and they get summarised by an LLM call to preserve context capacity.
- Roller - The dialectic machine
Roller works in Tandem with the Mipsy text archive, but it works like a theatric play: You define self-contained contexts (worlds) which contain âStoriesâ and âCharactersâ who talk to each other on a stage. You can either literally play this like a game, but obviously the characters can also be your research assistants. If you get salient âconvosâ, these can be dragged back into the Mipsy text store.
Iâve tested these on a Mac mini (16 GB M1), and a MacBook Pro with an M1 Max processor and 64 GB of RAM. The mini ran a 12b model which worked surprisingly well both in the language-producing tasks and for writing scripts. The MacBook I successfully tested on a 24b model and a 40k context in Roller.
You can download these apps either as stand-alone versions, or as an app that combines the two into one user-interface (Acopa).
I disclaim all responsibility for these apps. You download and run them on your own risk. If they donât work for you, ask a coding assistant to modify them.