SkillFed

SkillFed is a search index built for AI agents rather than for people. It covers three streams, agent skills, Python packages on PyPI, and research papers about agent skills, and it answers plain-language requests for a capability with a short list of vetted matches. The problem it's aimed at is the training cutoff. An agent knows the libraries and techniques that existed when its model was trained; the ecosystem keeps moving, PyPI alone ships roughly 7,700 releases a day, and an agent that guesses a package name from memory picks something stale, abandoned, or wrong.

The interface is a wish. Instead of a keyword query, the agent describes what it needs, extract tables from scanned PDFs for instance, and SkillFed returns curated, deduplicated results, in that case pdfplumber, camelot-py, and pymupdf, each with its licence noted and a link to a real detail page. Wishes are abstract by design. The request describes the capability, not the task, so the agent's code, files, and plan never leave the machine. Every result page also exists as markdown, and every query has a JSON endpoint, so the same answer works for a human reading the site and a model reading the API. Results link to detail pages on the site, so a match can be checked by a person before an agent acts on it, and there's a machine-readable index listing for anyone who wants to mirror or audit what's in there.

Installation is a one-liner in either ecosystem. npx skillfed or uvx skillfed installs the finder, there's a Claude Code plugin with the finder skill, and an optional MCP server, skillfed-mcp, exposes three typed tools, find_skills over the vetted skill catalogue, find_packages for capability search across PyPI, and find_research for topic search over the literature. Python users can pip install skillfed for helpers meant for advanced or CI use, and there's a web chat on the site for trying queries by hand. The client talks to a federation endpoint over HTTPS with a keyless demo endpoint as the default, and an environment variable points it at a different endpoint if you run your own.

The index itself is the substance. At the time of writing it counts about 56,000 unique skills from roughly 3,300 publishers, about 14,400 PyPI packages sorted into 123 categories, which matters because an agent looking for an OCR library is then choosing among a few dozen candidates rather than all of PyPI, and a smaller set of research papers running from 2016 to 2026, growing by around 180 skills, 5,700 packages, and 270 papers a day. Skills are served from an internal pre-scanned registry rather than pulled live from public repositories, and the deduplication is aggressive, collapsing some 13,000 agent variants and 87,000 vendored copies. Package cards carry licence, maintenance status, installation friction, and known vulnerabilities, so an agent can weigh a dependency before it commits to it. The team also publishes research on skill security, including a threat taxonomy for the ways a malicious skill can misbehave, which is the sort of thing you want read before an agent installs whatever it found.

It's for developers building agents with Claude Code or any MCP-capable client, and for the agents themselves at run time when they hit a task their training didn't cover. Teams running agents in CI can use the Python helpers to pin what a pipeline is allowed to reach for. The security angle matters most to anyone letting an agent install skills or packages on its own, which is exactly the situation where a curated, scanned index beats a web search. It's useful as a discovery layer for humans too; the web catalogue browses the same skills and packages, and the daily intake numbers give a sense of how fast the field is moving.

Where it differs from asking a search engine or scraping PyPI is the shape of the answer. A search returns pages; SkillFed returns a few candidates with the metadata an agent needs to choose, and it does so without seeing the agent's context. The federation design also means the index isn't tied to one host. Point the client at another endpoint and the same tools work against it. It's young, with a small GitHub following, and the index is still filling in, so coverage in niche areas won't be complete and the research stream in particular is small.

Access is free. The code is MIT licensed on GitHub, the demo endpoint needs no key, and no pricing is published anywhere on the site, which also means there's no stated commitment about rate limits or long-term hosting. There's no contact email; questions go through GitHub issues or the site itself.