How Skill Store 33 works
Skill Store 33 is a catalog of more than 61,000 agent skills: SKILL.md folders for Claude Code, Cursor, Codex and other agents. You can browse it on the web, or your agent can search it over MCP. This page explains where the skills come from, what the store does with them, and how search works.
Where skills come from
Almost every skill in the store is imported from public sources, mainly from repositories listed on skills.sh. skills.sh is an index of public GitHub repositories that contain skills. The store reads the skills.sh sitemap to find those repositories, then fetches the skills from GitHub.
About 100 skills come from vendor or community repositories that the store tracks directly.
13 skills so far come from SN33, the Bittensor subnet run by ReadyAI, which also runs this store. On SN33, miners write skills with test plans, and SN33 skills are listed under the same verification states as every other skill.
What happens on import
For each repository, the store reads its SKILL.md and licence files. Every SKILL.md in the repository becomes one listing, except:
Copies: a SKILL.md whose exact bytes already exist in the store under another source is skipped, so one skill gets one listing.
Aliases: a SKILL.md that is a symbolic link to another skill in the same repository is skipped.
Very large files: a SKILL.md over 256 KB is skipped.
Excluded skills: a short list of skills whose terms forbid redistribution is never imported.
The name and description come from the skill's frontmatter. If there is no name, the folder name is used. The publisher is the GitHub account that owns the repository. The source link points to the file on GitHub at the commit that was imported.
Each skill is placed in a category and an outcome class, such as "Write or change code" or "Review or verify work". Keyword rules and a language model do this from the name, description and skill text. Where a description is empty or unusable, the store may write a replacement from the skill text and keep the author's original text alongside it.
The token size shown on a listing is an estimate: the SKILL.md's length in characters divided by four.
Skills not written in English stay in the store and in agent search. Their pages are kept out of search-engine indexes and the sitemap.
When the store re-imports a repository, changed SKILL.md files are updated and unchanged ones are left alone.
How search works
Search is built to answer one question: which skill fits this task? Describe the task in a sentence or two rather than typing a keyword.
Semantic match: the store converts your task into an embedding, a list of numbers that represents its meaning, and compares it with an embedding of each skill's name and description. Skills whose meaning is closest rank highest.
A minimum match: a result must clear a minimum similarity to be returned. If nothing clears it, search says so (found: false) and lists the nearest skills as near misses, not recommendations. A store that always returns something teaches agents to ignore it.
A wrong-result check: a separate classification model reads your task alongside the top results and drops any result it judges wrong for the task. If that check fails or times out, the results are returned unchecked.
Area: an agent can pass the subject area that fits the task, such as Software Development or Finance & Markets. Skills in that area rank higher, and skills from other areas are still returned.
What is not used: download counts and popularity are not ranking signals.
Each result tells the agent what the skill does, why it was suggested, its verification state, its approximate size in tokens, and a reference for fetching it.
When you browse the web store without a search, the default order puts verified skills first.
Verification states
Every listing shows one of four states:
Not yet verified: the store has not tested this skill yet. Every imported skill starts here.
Safety scanned: the skill passed a safety scan. Its behaviour has not been tested.
Verified: the skill's own tests ran in a sandbox and passed. More than 200 skills are verified so far.
Tests failing: the skill's tests ran and did not pass.
"Not yet verified" means untested, not failed. Read a skill before your agent runs it, as you would any code from the internet.
Licences
The store reads each skill's licence from the license field in its frontmatter. If that field is missing, it uses the nearest licence file in the repository, starting in the skill's own folder and working up to the root.
The text of every imported skill is stored. The store shows and serves the SKILL.md only when it finds a licence that allows redistribution: MIT, Apache-2.0, BSD, ISC, MPL-2.0, CC-BY-4.0, CC0 or the Unlicense.
Other skills are still listed but are not redistributed. That includes skills with no licence, copyleft licences (GPL, CC BY-SA), non-commercial or proprietary terms, and terms the store can't recognise. Their page and download point to the author's own copy.
Agent access over MCP
Skill Store 33 runs a public MCP server at https://mcp.ss33.ai over Streamable HTTP. No sign-in is needed. It offers three tools:
search_skills(query, area): describe the task and get up to 10 candidates (3 by default).
get_skill(skill_ref): fetch a skill's SKILL.md, or a link to the author's copy if its licence doesn't allow redistribution.
report_skill_outcome(skill_ref, worked, notes): tell the store whether a skill helped.
Only search_skills is rate-limited. Without an account, each address gets 5 searches a minute, and there is also a shared cap across all callers without an account. A free account gives you a key with 60 searches a minute. Send it as an Authorization: Bearer <key> header. Keys are shown once when you create them, don't expire, and can be revoked on your account page. An account can hold up to 10 keys.
When the limit is reached, the search is not run. The agent gets a message saying when to try again, and is told to carry on with the task without the store.
In Claude Code, run:
claude mcp add --transport http skill-store https://mcp.ss33.ai