mcp-skill

# mcp-skill Turn any MCP server into a typed Python SDK. > **Compile MCP tools into code.** `mcp-skill` introspects an MCP server and generates a Python class where each tool becomes a typed async method. ## Before and After ### Before Agents call MCP tools through the model loop — one round-trip per tool call: ``` llm.call_tool("web_search_preview", {"query": "..."}) # → model decides next step → calls another tool → model decides again → ... ``` ### After Agents call tools directly in code: ```python result = await app.web_search_preview( objective="find latest news", search_queries=["topic X 2026"] ) # Agent processes result in code — no round-trip back to model ``` ## How It Works ``` ┌─────────────┐ ┌───────────────┐ ┌────────────────────┐ │ MCP Server │─────▶│ mcp-skill │─────▶│ Generated Skill │ │ (any URL) │ │ CLI │ │ │ │ │ │ │ │ app.py │ │ Tools: │ │ 1. Connect │ │ ├─ Typed class │ │ - search │ │ 2. Introspec │ │ ├─ Async methods │ │ - fetch │ │ 3. Map types │ │ ├─ Auth + storage │ │ - ... │ │ 4. Generate │ │ └─ JSON parsing │ │ │ │ 5. Validate │ │ │ └─────────────┘ └───────────────┘ │ SKILL.md │ │ └─ Agent docs │ └────────────────────┘ ``` 1. Connects to the MCP server using [fastmcp](https://github.com/jlowin/fastmcp) 2. Introspects all available tools via `list_tools()` 3. Converts each tool's JSON Schema into Python type annotations 4. Generates a typed `App` class where each MCP tool becomes an `async` method 5. Validates the output with `ast.parse` → `ruff` → `ty` 6. Generates `SKILL.md` with tool documentation and usage examples for agents ## Motivation MCP servers give agents access t…