mcp -workshop

This MCP server is a lightweight backend built with the TypeScript MCP SDK to expose tools (functions), memory, and structured prompts to an AI model like Claude or another MCP client. It acts as the brain behind an AI assistant or agent by providing: 1) tool access — register and expose callable tools that the model can invoke during a session; 2) memory and context management — store conversation state and memory across interactions; and 3) prompt orchestration — manage the prompts and responses that guide the model’s behavior. You can run this locally for development and testing, then deploy to Cloudflare Workers or Pages for a live, edge-hosted MCP server accessible on the web. The setup uses Node.js v18+ and the MCP SDK to wire everything together, making it straightforward to iterate on tools, memory schemas, and prompts as you build more capable AI backends.

To use it, start the local server and connect your MCP client (for example Claude or another LLM backend) to the exposed endpoints. You’ll register tools (functions) that the AI can call, implement memory storage for chats or state, and define prompts that shape how the model should respond in different contexts. The workshop flow covers deploying to Cloudflare so others can reach your MCP server via web requests, enabling real-world testing and demonstrations.

Prerequisites:\n- Node.js v18+ installed on your machine\n- npm (comes with Node.js)\n- Access to an MCP client or Claude account for testing the model backend\n- (Optional) Cloudflare account for deployment\n\nSteps:\n1) Prepare your environment\n- Ensure Node.js is installed: node -v (should show v18.x.x)\n- Create a new project directory for the MCP workshop server and navigate into it.\n\n2) Initialize and install dependencies\n- git clone <repository-url-for-workshop> (or create your own project structure)\n- cd into the project directory\n- npm install\n\n3) Implement or adjust the server entry point\n- Ensure there is a server.js (or equivalent) that boots the MCP server using the MCP SDK.\n- If using TypeScript, build steps may be required (e.g., npm run build) to produce the JavaScript output expected by node.\n\n4) Configure environment variables\n- PORT: port to run the server on (default if not provided by code)\n- CLAUDE_API_KEY: your Claude API key if you’re using Claude as the model backend.\n\n5) Run the server locally\n- npm start or node path/to/server.js\n- Verify the server starts and listens on the configured port.\n\n6) Optional — deploy to Cloudflare\n- Follow Cloudflare Workers/Pages deployment instructions in the workshop materials to publish your MCP server to the edge. Ensure the deployed URL is reachable and that the model client can access it.

Tips and common notes:\n- Ensure you have a valid API key for your chosen MCP client (e.g., Claude) and that network requests from the server to the model backend are allowed by your environment.\n- When exposing tools, document their inputs/outputs clearly so the AI can invoke them reliably.\n- For memory, define a simple schema (e.g., per-conversation chat history) and consider persistence (in-memory vs. database) depending on the workshop scope.\n- If deploying to Cloudflare, you may need to adapt the server to run as a Worker-compatible module or use Pages Functions.\n- Check your environment variables on the hosting platform and keep sensitive keys secure (do not commit to source).\n- Common issues include port conflicts, network egress restrictions, and API rate limits from the MCP model provider.