Skyll provides a hosted MCP server at api.skyll.app/mcp, exposing a set of MCP tools that let agents discover, fetch, and learn skills on demand without pre-installation. The hosted server offers the following tools: search_skills to query for relevant skills using natural language, add_skill to retrieve a skill by name or by full path, get_skill to fetch a specific skill by source and id, and get_cache_stats to inspect cache utilization. The add_skill tool is designed to be the simplest way for agents to learn skills, returning the best-matching skill for a given name or the explicit skill when a full path is provided. When self-hosting Skyll, you run your own MCP server that points to the same toolset so agents can operate entirely within your environment. The MCP interface mirrors common MCP clients, making it easy to integrate Skyll with Claude Desktop, Cursor, or other MCP-enabled tools.

Prerequisites:

• Python 3.8+ and pip
• Git
• (Optional) Virtual environment support (venv) for isolation

Self-hosted installation steps:

1. Clone the repository git clone https://github.com/assafelovic/skyll.git cd skyll

2. Install the package in editable mode including server dependencies pip install -e ".[server]"

3. (Optional) Configure a GitHub token for higher rate limits echo "GITHUB_TOKEN=your_token_here" > .env

4. Start the MCP server (self-hosted) uvicorn src.main:app --port 8000

   or, if using the provided module as an MCP server:

   python -m src.mcp_server --transport http --port 8080

5. Test the server locally curl "http://localhost:8000/search?q=react+performance&limit=5"

Configuration tips:

• Environment variables: GITHUB_TOKEN for higher GitHub rate limits; CACHE_TTL to control cache duration (default 86400 seconds); ENABLE_REGISTRY to toggle the community registry.
• If hosting behind a reverse proxy, ensure proper TLS termination and port exposure for the chosen transport (http, sse).
• For self-hosted MCP usage, you can point clients to the local endpoint (e.g., http://localhost:8080) or the externally exposed URL if deployed behind a gateway.
• The hosted MCP configuration example uses a URL-based approach, while the self-hosted setup demonstrates a Python-based MCP server entry point. Pick the approach that matches your deployment.
• When upgrading Skyll, re-check the MCP endpoint behavior and ensure compatibility with your MCP clients.