Install Skillibary in your agent

This page covers a single goal: get an MCP-compatible agent (Claude Desktop, Cursor, Continue) talking to the public Skillibary registry. After you finish, the agent gains four tools and can search and load any verified skill on demand.

What gets installed

Four tools become available to the agent once the server is wired up:

Submitting and re-verifying skills is intentionally not exposed through MCP. Those operations require your GitHub account, which the agent doesn't have. Use the web submit form or the CLI's publish command for that flow.

Step 1 — Get your API key

Sign in with GitHub, then open Settings → API Key and click “Generate API Key”. Copy the key — it's shown only once. It looks like sk-xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx.

You can rotate or revoke it from the same settings panel at any time. One key per account.

Step 2 — Add the config

Paste this block into your agent's MCP config, replacing sk-your-key-here with the key you just copied:

{
  "mcpServers": {
    "skillibary": {
      "command": "npx",
      "args": ["-y", "skillibary-mcp"],
      "env": {
        "SKILLIBARY_API_KEY": "sk-your-key-here"
      }
    }
  }
}

The -y flag tells npx to install the package without an interactive prompt. Without it, the first run can hang waiting for confirmation that the agent never sees.

If you only want browsing (no semantic search), the key is optional:

{
  "mcpServers": {
    "skillibary": {
      "command": "npx",
      "args": ["-y", "skillibary-mcp"]
    }
  }
}

Step 3 (optional) — Add a CLAUDE.md hint

By default, Claude will call Skillibary tools whenever it thinks they might help. If you want more control — for example, only use skills when you explicitly ask — add the following snippet to your CLAUDE.md (or .github/copilot-instructions.md / system prompt file):

## Skillibary MCP

You have access to the Skillibary registry via four MCP tools.

Call `search_skills` or `list_skills` **only** when the user explicitly
asks to find, browse, or use a skill from the registry, or when you are
confident a published skill would directly improve the answer.

Do NOT call Skillibary tools for general coding questions, explanations,
or tasks you can handle with your built-in knowledge. Registry lookups
add latency — use them when the value is clear.

Typical flow:
1. search_skills — find candidates by semantic query
2. get_skill — load the full SKILL.md for the best match
3. Apply the skill instructions to the user's task

Adjust the wording to match your workflow. The goal is to give the model a clear signal about when the overhead of a registry lookup is worth it.

Where each client stores its config

Claude Desktop

If the file doesn't exist yet, create it with a single empty object {} and paste the mcpServers key inside.

Cursor

Open Settings, navigate to Features → MCP, click “Add MCP Server”, and paste the skillibary block (just the inner object, not the outer mcpServers wrapper). Cursor restarts the server automatically when you save.

Continue (VS Code)

Edit ~/.continue/config.json. The mcpServers key lives at the top level of the config object, with the same shape shown above. Reload the VS Code window to pick up the change.

Verifying the install

Open a new chat in your client and ask:

What MCP tools do you have available right now?

You should see four tools listed with names prefixed by skillibary. If the agent says it has no MCP tools, or lists tools from other servers but not Skillibary, jump to the troubleshooting section below.

Example agent prompts

Once the tools are present, the agent picks them up on its own:

Find a Skillibary skill that helps write good git commit messages.

Use the skillibary "pdf-extract" skill to pull text from this attached PDF.

List the top three Skillibary skills tagged "sql" by use count.

Two calls happen under the hood. First the agent invokes search_skills with the query, reads the summaries, and picks one. Then it calls get_skill with the id from the search result. The skill body lands in the agent's context, the use counter goes up by one, and the agent answers from the skill text.

Troubleshooting

The agent doesn't see any skillibary tools

1. You refreshed instead of restarted. Claude Desktop keeps the old subprocess running. Fully quit and relaunch.
2. JSON syntax error. A trailing comma or missing quote in the config silently disables the whole mcpServers block. Run cat claude_desktop_config.json | jq. If jq rejects it, your client did too.
3. npx is not on the agent's PATH. Common on macOS when Node was installed through a version manager (nvm, fnm, asdf). Easiest fix: replace "command": "npx" with the absolute path returned by which npx in your shell.

Tools listed, but every call returns an error

The MCP server prints diagnostic output to stderr; your client's log panel shows it.

The agent picks the wrong skill

Semantic search returns the closest match by cosine similarity, not by exact name. Two ways to help:

• Be specific in the prompt. “Find a skill that summarises PDFs into bullet points” beats “help me with this PDF”.
• Mention the slug. If you already know which skill you want, name it: “Use the skillibary ‘pdf-extract’ skill…”. The agent will search for that name, find it as the top hit, and load it directly.

Uninstalling

Remove the skillibary entry from your client's MCP config and restart the client. The npx cache will still hold the package, but it stays inert until the config references it. To purge the cache entirely: npx clear-npx-cache.