For the next model training version, would it make sense to incorporate all of these in the base model?

This is nothing like Dockerhub and, I'm sorry, but it's seriously useless. In its current state its worse than basically anything else.

You have no versioning, no automated or simplified update, no way to verify the authors, etc. The "installation" is literally just a wget.

This is a really poor solution for the moment, and honestly I think for the forseable future. I don't see how anything beyond git is necessary for skills management.

Most of the skills currently hosted are also really bad. They are just a duplicate of the information that MCP would give the models.

I do like the idea of crowd-sourced collections of resources like skills.

It might be more useful if it was an index of skills managed in GitHub. Sort of like GitHub actions which can be browsed in the marketplace[1] but are ultimately just normal git repos.

[1] https://github.com/marketplace?type=actions

I don't understand how "agent-browser" works.

Is it just the instructions? Where is the browsing executed? Locally with pupetter? Or it uses some service?

AI agent skills are very useful. Unlike MCP they do not waste context. Most of the time I am building skills that are very particular to my project. But occasionally I do use a skill that is more generic. Particularly when something is too new to have made it into the LLM training data set. Or not common enough.