?
What is MCP?
The Model Context Protocol is the open standard for connecting AI assistants to external tools and data. An MCP server exposes named tools (and optionally resources and prompts) that clients like Claude Desktop, Claude Code, Cursor, and Windsurf can call. One server works in any MCP-compatible client.
?
Do these actually run, or are they just examples?
They're complete, runnable servers — not pseudocode. Every template has install + run steps, a working entrypoint, dependencies pinned, and the exact client config snippet to wire it up. The TypeScript templates are tested to compile cleanly against the SDK; the Python templates' safety layers (SQL validator, AST safe-eval, JSON path walker) are exercised before shipping.
?
TypeScript or Python?
Both. The pack ships five TypeScript templates (using @modelcontextprotocol/sdk) and five Python templates (using FastMCP from the official mcp package). The hello-world is in both languages side-by-side so you can see the same shape twice.
?
Which MCP clients are supported?
Any MCP-compatible client. Setup snippets and a config generator are included for Claude Desktop, Cursor, Windsurf, and Claude Code. For other clients (or for connecting the OAuth server via HTTP), the protocol is the same — the JSON-RPC and transports are spec-defined.
?
Can I use these commercially?
Yes — unlimited personal, internal, and client/work use. Build commercial products on top, ship them under your own name. The only restriction is that you can't resell the pack itself (or the templates substantially unchanged) as a competing starter pack. See LICENSE.txt for the exact terms.
?
What versions of the SDKs do these target?
TypeScript templates target @modelcontextprotocol/sdk ^1.29 with Node.js 18+. Python templates target the official mcp package (>= 1.6, includes FastMCP) on Python 3.10+. Both use the modern stable APIs (McpServer + registerTool in TS; @mcp.tool() in Python).
?
Is the RAG template tied to OpenAI?
No — it has a pluggable embeddings layer. You can use OpenAI embeddings (with an API key) or local sentence-transformers (no API key, downloads a ~90 MB model on first run). The vector store is an on-disk JSON file with documented swap-in points for Qdrant, pgvector, or Milvus.