← All Tracks

MCP Integration

Master the Model Context Protocol — from understanding the spec to building servers and orchestrating multi-agent systems.

intermediate 4 stages 15 hours not started

Prerequisites: Agentic Workflows

Claude CodeNode.jsMCP SDK

Stage 1: MCP Concepts — What and Why

You’ll know this when… you can explain what MCP does, why it exists, and how it differs from regular API integrations or tool-use — in plain English.

Key Concepts

  • MCP = a standard protocol for connecting AI models to external tools and data
  • Client-server architecture: the AI app is the client, your tools are servers
  • Resources (data the model can read), Tools (actions the model can take), Prompts (reusable templates)
  • Why standardization matters — write once, use with any MCP-compatible AI app
  • How Claude Code already uses MCP under the hood

Practice Project

Inventory your existing integrations. List every external service your Intelligence Hub touches (GitHub API, RSS feeds, Claude API). For each one, write a paragraph on how it could be exposed as an MCP server. Which would benefit most from standardization?

Stage 2: Your First MCP Server

You’ll know this when… you’ve built a working MCP server that exposes at least one tool and one resource, and connected it to Claude Code.

Key Concepts

  • Setting up the MCP TypeScript SDK
  • Defining tools with input schemas (JSON Schema)
  • Exposing resources (static data or dynamic queries)
  • Transport layer: stdio (local) vs. HTTP/SSE (remote)
  • Testing your server with the MCP Inspector

Practice Project

Build a “briefing search” MCP server. Create an MCP server that exposes your Intelligence Hub briefings as a searchable resource. Tools: search_briefings(query) returns matching items across all briefings, get_briefing(date) returns a specific day’s briefing. Connect it to Claude Code so you can ask “what did the briefing say about MCP last week?” in conversation.

Stage 3: Connecting to Tools — Real-World Integrations

You’ll know this when… you can build MCP servers that wrap real APIs and services, handle authentication, manage rate limits, and return useful structured data.

Key Concepts

  • Wrapping REST APIs as MCP tools
  • Authentication patterns (API keys, OAuth tokens via environment variables)
  • Error handling and graceful degradation
  • Rate limiting and caching strategies
  • Composing multiple tools in a single server

Practice Project

Build a “site manager” MCP server. Wrap your existing GitHub API publishing workflow (agent/publish.js) as an MCP server with tools: publish_content(path, content), list_recent_deploys() (via Vercel API), check_build_status(). Now you can manage your site from any MCP-compatible AI app.

Stage 4: Multi-Agent Orchestration

You’ll know this when… you can design a system where multiple agents coordinate via MCP, each with its own tools and responsibilities, working toward a shared goal.

Key Concepts

  • Agent roles: researcher, analyst, writer, publisher
  • MCP as the communication layer between agents
  • Shared state via MCP resources
  • Orchestration patterns: sequential handoff, parallel fan-out, supervisor
  • Debugging multi-agent systems — tracing requests across agents

Practice Project

Refactor the Intelligence Hub as a multi-agent system. Design (and optionally build) a version where: Agent 1 (Researcher) handles ingestion and dedup, Agent 2 (Analyst) does evaluation and scoring, Agent 3 (Writer) generates the briefing, Agent 4 (Publisher) handles GitHub/Vercel deployment. Each agent is an MCP server. The orchestrator calls them in sequence. Compare the architecture to your current monolithic run.js.