Documentation
Skillibary is an open registry of agent skills. Most people land here for one of two reasons: they want to plug an existing agent into the registry, or they want to author a skill and ship it. Pick the matching section below.
The full set of guides lives in this section. The repo also has plain Markdown versions of each page under docs/, which is what you read on GitHub.
A one-time setup that gives your agent six tools: search the registry, load any skill on demand, log usage. Includes the exact config file paths per client and the failure modes you will hit.
Every field the registry parses, every limit the server enforces, what gets stripped on submit, and a worked example you can adapt.
install, validate, and publish from the command line. Hooks into the same registry the web UI uses.
The six public HTTP endpoints, with exact request and response shapes, rate limits, and the error codes you can actually hit.
What this registry is
A skill is a Markdown file with YAML front-matter. The body documents what the skill does; the front-matter is the machine-readable handle: a slug name, a one-line description, a version, and tags. Authors write these locally and submit through the web form or the CLI. Every submission is auto-verified by Claude before it can appear in search results.
On the consumer side, the registry exposes an MCP server. Drop it into Claude Desktop and your agent can call search_skills over a vector index, then get_skill to load the full body of whatever it picked. The server is generic; it does not bundle a specific skill subset.
What this registry is not
It is not a package manager. Skills are read by the agent at runtime, not compiled or executed. The CLI's install command only writes a SKILL.md to disk so you can edit or version-control it; the actual execution model is whatever your agent decides to do with the text.
It is also not a marketplace. Everything is MIT-licensed, free, and moderated by automated verification plus the occasional human pass. Skills tagged with anything resembling malicious intent are rejected at submission.