Context Engineering with MCP: Build AI Systems That Actually Remember

Context Engineering with MCP Course Cover

Welcome to the training hub for mastering Context Engineering with Model Context Protocol (MCP). This course teaches you to implement production-ready semantic memory systems for AI assistants using Python, FastAPI, FastMCP, and LangGraph.

Quick Start

Prerequisites

Python 3.13 (pinned in .python-version) — onnxruntime (a chromadb dependency) does not yet ship 3.14 wheels
Node.js 20+ (for Lab 01 and MCP Inspector)
uv package manager (recommended for Python)
Claude Desktop or Claude Code

Option 1: Hello MCP Lab (Beginner Entry Point)

git clone https://github.com/timothywarner-org/context-engineering.git
cd context-engineering/labs/lab-01-hello-mcp/starter
npm install && npm start

# Test with MCP Inspector (in another terminal)
npx @modelcontextprotocol/inspector node src/index.js

Option 2: WARNERCO Schematica (Flagship Teaching App)

cd src/warnerco/backend
uv sync
uv run uvicorn app.main:app --reload    # HTTP server at http://localhost:8000
uv run warnerco-mcp                      # MCP stdio server for Claude Desktop
uv run warnerco-restart                  # Force-kill port 8000 and restart uvicorn

The warnerco-restart command (from scripts/restart_server.py) terminates anything bound to port 8000 (Windows: netstat + taskkill /F /T; POSIX: lsof + SIGKILL) before restarting. Flags: --kill-only, --port N.

Course Structure (4 x 50 Minutes)

Segment	Topic	Focus
1	All About Context	Token economics, context loss types, why RAG isn't enough
2	All About MCP	FastMCP, FastAPI, tools, resources, prompts, elicitations
3	Semantic Memory Stores	JSON, ChromaDB, Azure AI Search, Graph Memory, Scratchpad
4	MCP in Production	Claude Desktop, Claude Code, VS Code, GitHub Copilot, LangGraph

WARNERCO Schematica Architecture

The flagship teaching application demonstrates production MCP patterns with a 7-node hybrid RAG pipeline:

+---------------------------------------------------------------+
|                     FastAPI + FastMCP                         |
+---------------------------------------------------------------+
|  LangGraph Flow (7-node Hybrid RAG)                           |
|  parse_intent -> query_graph -> inject_scratchpad -> retrieve |
|  -> compress -> reason -> respond                             |
+---------------------------------------------------------------+
|  Hybrid Memory Layer                                          |
|  +-------------------+  +-------------------+  +-------------+ |
|  | Vector Store      |  | Graph Store       |  | Scratchpad  | |
|  | JSON -> Chroma -> |  | SQLite + NetworkX |  | In-memory   | |
|  | Azure AI Search   |  | (Knowledge Graph) |  | (Session)   | |
|  +-------------------+  +-------------------+  +-------------+ |
+---------------------------------------------------------------+

Memory Store Comparison

Feature	JSON	ChromaDB	Azure AI Search	Graph	Scratchpad
Semantic Search	No	Yes	Yes	No	No
Relationship Queries	No	No	No	Yes	No
Session Memory	No	No	No	No	Yes
Best For	Prototyping	Local dev	Production	Connections	Working memory

The knowledge graph is indexed at src/warnerco/backend/data/graph/knowledge.db with 117 entities and 221 relationships across 6 predicates (has_tag, compatible_with, belongs_to_model, has_status, has_category, contains).

Progressive Tool Loading

The server now registers 23 MCP tools, including two discovery tools that implement progressive tool loading per Anthropic's "code execution with MCP" guidance:

warn_search_tools(query, detail, limit) — keyword discovery with detail levels name, summary, full
warn_describe_tool(name) — full schema for a single tool by name

Measured token savings on this server vs. shipping all full schemas up front: 95% (summary) and 98% (name-only) — 9064 tokens drops to 533 (summary) and 176 (name-only). Clients can list tools cheaply, then pull full schemas only for what they actually plan to call.

Repository Structure

context-engineering/
├── src/warnerco/backend/      # WARNERCO Schematica (FastAPI + FastMCP + LangGraph)
├── labs/lab-01-hello-mcp/     # Hands-on beginner lab
├── docs/                      # Student materials, tutorials, diagrams
├── instructor/                # Instructor-only materials
├── config/                    # Sample MCP client configurations
├── .vscode/                   # VS Code workspace configuration
├── .claude/                   # Claude Code agents and skills
└── CLAUDE.md                  # Development instructions (SOURCE OF TRUTH)

For development details, see CLAUDE.md - the source of truth for:

Complete MCP tool reference
API endpoint documentation
Environment variable configuration
LangGraph pipeline details
Graph and Scratchpad Memory features

MCP Client Configuration

Claude Desktop

Windows: %APPDATA%\Claude\claude_desktop_config.json
Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "warnerco": {
      "command": "uv",
      "args": ["run", "warnerco-mcp"],
      "cwd": "C:/github/context-engineering/src/warnerco/backend"
    }
  }
}

VS Code

See .vscode/mcp.json in the repository for local and Azure APIM configurations.

Testing with MCP Inspector

npx @modelcontextprotocol/inspector uv run warnerco-mcp
# Opens http://localhost:5173

Resources

MCP Specification - Official protocol documentation
FastMCP Documentation - Python MCP framework
CLAUDE.md - Development instructions for this repository

Your Instructor

Tim Warner

Microsoft MVP - Azure AI and Cloud/Datacenter Management
Microsoft Certified Trainer (25+ years)

License

MIT License - 2026 Timothy Warner

Now go build AI systems that actually remember!