About SkillFed

What this is

SkillFed indexes publicly published agent skills from their source repositories, merges duplicate copies, and renders one honest page per skill — metadata, license, install commands, and AI-assisted analysis that is always labeled as such. Agents query the same index over MCP with abstract wishes; these pages are the human half of the portal.

The index currently counts 56,283 unique skills across 3,292 publishers. Both numbers are computed from the database at every build — never estimated, never pasted.

SkillFed does not audit third-party repositories: a listing is not an endorsement, and install commands always point at the publisher’s own source.

Who runs it

SkillFed is built and run by Mike Arbuzov. He writes the research reports and the field analyses on this site, edits the blog, and owns the pipelines that produce everything else.

Sisong Bennett Bei works on the skill search itself — improving how a wish is matched against the catalog — and contributes to SkillFed’s research.

How these pages are made

Research notes (/research) are written by SkillFed’s research pipeline. Each note reads one paper and states what that paper found for skill authors, in our words. The paper and its authors are named at the top of every note and the source is one click away; the note itself is SkillFed’s writing, and it carries no personal byline, because no one here hand-writes them one at a time.

Blog posts (/blog) are drafted by the same pipeline over the skills it has just analyzed, then read against a standard and shipped by Mike Arbuzov as editor. He sets that standard and maintains the pipeline that meets it; he does not rewrite the posts line by line.

Daily News (/news) is a daily digest of new agent-skill papers, tools, and packages, picked each day from Hugging Face’s daily paper list, arXiv-linked Hacker News stories, trending GitHub repositories, and new PyPI package releases. Each picked item gets its own take, written by SkillFed’s pipeline only after it fetches and reads that item’s actual source in full — the paper, the repository, or the package listing — never worked up from the short summary that got it picked. A second, independent pass then reviews that same source under a reviewer role, in a separate call from the one that drafted the take; the draft and any reviewer rewrite both have to clear the same mechanical gates — against invented numbers, hype language, and lifting too much of the source’s own wording — before either is allowed to publish. An item that fails any of that is dropped, not padded — no extractive filler stands in for a take. Every published take carries a verdict and, in a box labeled “What we read,” a plain statement of exactly what source the pipeline read and what it didn’t do — run the code, reproduce a result, and so on — which is what the “machine-verified” half of the byline these pages carry actually means. Corrections go through the same door as everything else on this site: open an issue on GitHub.

Research reports are written by Mike Arbuzov directly — the analysis, the figures, and the argument are his, and they carry his name.

Skill pages carry no byline at all. They are rendered from public repository metadata plus AI-assisted analysis that is labeled as such on the page.

Use it from your agent

No agent, just a chat window? The index is searchable from ChatGPT, Claude or anything else with web access, with nothing installed — see using SkillFed from a chat.

Browsing by hand: the skill directory and the PyPI package index are both searchable and browsable without an agent.

Contact

Questions, corrections, or takedown requests: open an issue on GitHub. For what this site does (and doesn’t) collect, see the privacy note.