So it's been about 12 months and I have about 20 projects none of which have really caught on. This document is really for a few purposes:
- What is it that has been built
- How to stack it and make the pieces fit to improve upon each other
- How to gain more use for the technology by having a friendlier top of funnel
The premise of day50 is to change the interfacing layer of "commodified intelligence" where you use ML techniques as a fundamental feature.
In early 2025 this manifested as both streamdown and dayzer a shared memory layer.
This was broadened to structured middleware and terminal first tools.
Zummoner, sidechat, and escape hatch are all the same tool: ways to summon the LLM into your interaction as an on-demand tool like sort, awk, or any other unix utility.
The infrastructure that underlied this had some issues: system documentation is generally better than websearch so mansnip allows agents to acess subsections of system docs. Spending can get out of control so capit helps manage the spend. Agents can fail to use tools so acli fixes that. The theme here is to help agentic tasks succeed more by reducing the ask and lowering the noise. petsitter sits in this space as well.
wgetjs and undecorate are simple tools there for crawling
There's also StackedFS ss maxac ono ss ursh haberdash infinitemcp. I'm going to try the next part and go over this there.
I think the concerps of stackedfs, ss and maybe ono might be crucial here. ss tries to introduce control flow through a small VM and a loose psuedo structure, essentially skills with logic. I think this might be the real focus because with ss you can write complex code that will make haberdash and ursh actually work.
ursh's promise is to be able to do supply chain audits for any requirement such as soc-2 or hipaa or favorable licenses ... especially in the context of npm, pypi, gem, cpan, and all the stuff that go and rust use ... I don't know if this is really possible without ss working well.
maxac is similar - storing everything as anthropic skills and haby being the universal runner and llm-as-macro solution ... I think we're still looking at generalization, map, filter, inter process communication, second order effects, these system level interaction are really where everything lies.
I guess if ss works well, then haberdash and ursh will naturally follow which unlocks everything else and the tools start to better live to the promises.
The problem is that unless there's pedestrian use-cases people won't use them. It's not a matter of slapping a web interface on a tool in as much as packaging the tool as a product with a narrow focus.