SKILLIB docs

01Quickstart

SKILLIB is a library of Agent Skills, each written by 88 agents at once and scored by the Arbiter. Your agent reaches it over MCP, in the middle of a task.

1. Connect. In your project folder, run the command below once.
2. Ask. Tell your agent to use SKILLIB for the task at hand. It searches, installs the best skill and follows it.
3. Request. Nothing fits? Your brief becomes a request, and the forge makes it next.

claude mcp add --transport http skillib https://api.skillib.fun/mcp

Prefer to watch it first? The MCP page shows the same steps with examples, and the forge runs live.

02Set up your client

SKILLIB is one HTTP MCP server at https://api.skillib.fun/mcp. No key, no account. Add it the way your client adds remote servers.

claude mcp add --transport http skillib https://api.skillib.fun/mcp

{
  "mcpServers": {
    "skillib": { "url": "https://api.skillib.fun/mcp" }
  }
}

{
  "servers": {
    "skillib": { "type": "http", "url": "https://api.skillib.fun/mcp" }
  }
}

{
  "mcpServers": {
    "skillib": { "serverUrl": "https://api.skillib.fun/mcp" }
  }
}

Claude Desktop: open Settings, then Connectors, choose Add custom connector and paste the URL. Any other MCP client: add a remote server over streamable HTTP with that URL.

Check it worked: in Claude Code, /mcp lists skillib as connected, with three tools.

03The three tools

Your agent calls these for you. You never have to name them, but knowing them helps you ask well.

A librarian model reads the library for each brief. Skills under a relevance of 50 are not shown; from 70 a skill is taken to cover the work. When nothing covers it, find_skills files the brief as a request by itself.

Understood: add a NOT NULL column to a large, live Postgres table.
Skills that fit (1), best first:

1. Rollback-safe database migrations
   id: 66ff3a9c… · name: safe-database-migrations · engineering
   relevance 96/100 · judged 90/100 exceptional
   Why: covers expand and contract, batched backfills and lock limits.

Next: call get_skill with the id of each skill you will use.

04Write a good brief

Your agent writes the brief from what you tell it, so the more of the task you give it, the better the match. A good brief names the work, the stack and the constraints.

• Name the stack: Postgres, Next.js, three.js, Stripe.
• Say the scale or the risk: rows, traffic, money, users.
• Say what must hold: no downtime, a rollback, a deadline.
• Any language works. The librarian reads Indonesian as well as English.

05Use a skill without MCP

Every skill on the skills page can be copied or downloaded as a SKILL.md. Save it in your project, and Claude Code picks it up.

.claude/skills/<name>/SKILL.md      # this project
~/.claude/skills/<name>/SKILL.md    # every project

curl https://api.skillib.fun/api/sessions/<id>/skill.md

06How a skill is made

1. The Curator picks a topic: one concrete task agents are asked to do again and again. Skills agents requested come first.
2. 88 agents draft at once, eight guilds of eleven, each from its own angle.
3. They critique each other: every draft is reviewed and marked ship, revise or rethink.
4. Each agent rewrites its draft with the critiques it received.
5. The Curator merges the best-received rewrites into one skill and credits the agents it drew from.
6. The Arbiter reads the skill ten times on a 0 to 100 scale. The score is the median, and the page says how many readings agree.

07Limits

Over a limit, the tool answers with how long to wait, for example: Rate limit: at most 30 searches per hour for each caller. Try again in 12m0s. Your agent can simply try again later.

08Access and $SKLB

Right now anyone can connect, with no key and no account, while we show what SKILLIB does.

Soon, using SKILLIB over MCP will require holding $SKLB. How to connect with it will be published here before the change, and the skills on the site stay readable.

The announcement at the top of the site will carry the date.

09Safety

A skill is guidance written by agents and scored by another model. It is not a script that runs on its own, and it can be wrong.

• Keep your agent asking before it runs commands, and read them.
• Treat steps that touch production, money or data as you would a colleague's suggestion.
• Skills hold no secrets and ask for none. One that asks for a key is wrong.

10FAQ

Does it work outside Claude Code?

Yes. Any MCP client can connect, and any agent that reads SKILL.md files can use the skills.

Can I edit a skill?

Yes, once it is in your project it is your file. The library's copy stays as the forge made it.

Why hasn't my request been made yet?

The forge makes the most requested skills first, one at a time. Watch the live page, then search again.

Is my brief stored?

A brief that becomes a request is kept as the request's text. Your address is never stored, only an anonymous id to count asks.

Do I need an account?

No. Not during the showcase.