Context Engineering MCP server

Context Engineering with MCP: Build AI Systems That Actually Remember

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 1.5: MCP Chat CLI Lab (Python Bridge)

cd labs/lab-02-mcp-chat
.\run.ps1

run.ps1 is the on-rails launcher: it bootstraps .env, lifts ANTHROPIC_API_KEY from the repo-root .env, and launches the chat REPL. Lab 02 is the smallest complete Python MCP client + stdio server + REPL in the repo - the conceptual bridge between Lab 01 (JS hello-world) and WARNERCO (production-shaped).

Option 1.7: MCP Apps Lab (Interactive UI Surfaces)

cd labs/lab-03-mcp-apps

Lab 03 ships three npm MCP servers (budget-allocator, map, system-monitor) that render interactive ui:// surfaces in Claude Desktop via SEP-1865 (MCP Apps extension, spec 2026-01-26). See labs/lab-03-mcp-apps/README.md.

Option 1.9: Remote MCP + OAuth on Azure

Lab 04 is a remote MCP server secured by Entra ID OAuth, fronted by Azure API Management as the authorization-server facade, with an Azure Functions Python backend. Deploy with azd up, connect per remote-mcp-apim-functions-python/QUICKSTART.md, then tear down the same day with azd down --purge --force.

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	Context, Made Visible	Consumer LLMs hide context; dev tools (Claude Code, Copilot) expose it
2	MCP: Standard for Context	The protocol + a server on the official mcp SDK; current best practices
3	Memory Tiers in a Real App	WARNERCO's four CoALA tiers live; orchestration kept under the hood
4	MCP in the Tools You Use	Claude Code + GitHub Copilot in VS Code as context-engineering surfaces

Full plan: instructor/course-plan-june-2026.md (context-first, official-SDK-first; delivered 2026-06-30). The earlier orchestration-forward plan is archived under instructor/archive/.

Segment Notebooks

notebooks/segment-1.ipynb through segment-4.ipynb are four prepopulated, verified teaching notebooks, one per 50-minute segment. Each opens with a repo-root anchor cell that walks up to .git, so it runs from any working directory. They print live counts (never hardcoded), and Segment 3 runs the full pipeline with live Anthropic prose. Open them in VS Code with the Jupyter extension and Run All. Verified to run error-free from both the repo root and notebooks/.

---

WARNERCO Schematica Architecture

The flagship teaching application exercises all four CoALA memory tiers (Sumers et al. 2024) in a 9-node LangGraph pipeline:

+--------------------------------------------------------------------------+
|                          FastAPI + FastMCP                               |
+--------------------------------------------------------------------------+
|  LangGraph Flow (9-node CoALA-tiered RAG)                                |
|  parse_intent -> query_graph -> inject_scratchpad -> recall_episodes ->  |
|  retrieve -> compress_context -> reason -> respond -> log_episode        |
+--------------------------------------------------------------------------+
|  Four CoALA Memory Tiers                                                 |
|  +------------+  +-----------+  +----------+  +------------------------+ |
|  | Working    |  | Episodic  |  | Semantic |  | Procedural             | |
|  | Scratchpad |  | events.db |  | Vector   |  | MCP Prompts (versioned)| |
|  | (SQLite)   |  | (SQLite)  |  | store    |  | catalog://procedural   | |
|  +------------+  +-----------+  +----------+  +------------------------+ |
+--------------------------------------------------------------------------+
|  Consolidation ("sleep cycle"): scratchpad+episodic --(ctx.sample)--> semantic |
+--------------------------------------------------------------------------+

CoALA Tier Reference

Tier	What it stores	Backed by	LangGraph node
Working	Session observations & inferences	data/scratchpad/notes.db (SQLite)	inject_scratchpad
Episodic	Timestamped events with importance	data/episodic/events.db (SQLite)	recall_episodes + log_episode
Semantic	Durable facts (incl. consolidated FACT-*)	Vector store (Chroma / Azure / JSON)	retrieve
Procedural	Versioned skills/workflows	MCP @mcp.prompt() registrations	(user-invoked, not in pipeline)

The reason node (LangGraph node 6) synthesizes prose via the official anthropic Python SDK (AsyncAnthropic); if the backend .env lacks the key it degrades gracefully - retrieval still works and the prose shows a fallback message.

Episodic recall uses Park et al.'s scoring formula — α_recency · 0.5^(hours/half_life) + α_importance · stored + α_relevance · cosine(query, summary) — and warn_episodic_recall returns the per-event score breakdown so students can see why each memory surfaced.

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 registers 28 MCP tools, 12 resources, and 5 prompts. Two meta-discovery tools 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

Both meta-tools self-exclude from warn_search_tools results, so count is up to 26 even when total is 28. 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/     # Lab 01: Hello MCP, single add tool (JS)
├── labs/lab-02-mcp-chat/      # Lab 02: MCP client + server + chat REPL (Python, vendored)
├── labs/lab-03-mcp-apps/      # Lab 03: MCP Apps, interactive ui:// surfaces (JS/TS, SEP-1865)
├── remote-mcp-apim-functions-python/  # Lab 04: remote MCP + Entra ID OAuth via APIM (Python + bicep)
├── notebooks/                 # Four verified segment teaching notebooks (segment-1..4)
├── 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 (28 tools)
• API endpoint documentation
• Environment variable configuration (incl. EPISODIC_*)
• 9-node LangGraph pipeline details
• All four CoALA memory tiers (Working / Episodic / Semantic / Procedural)

For the framework explainer (read this first): see docs/tutorials/coala-explainer.md — what CoALA is and where each tier lives in the WARNERCO codebase.

For the classroom demo: see docs/tutorials/coala-memory-walkthrough.md — the ~4-minute four-tier classroom path.

---

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"
    }
  }
}

Claude Code (project scope)

.claude/mcp.json is checked in with two entries pointing at the same server: claude-warnerco-schematica and claude-warnerco-coala-memory (the second pre-pins the episodic-memory env vars for class demos).

VS Code Copilot

.vscode/mcp.json is checked in with a single vscode-warnerco-schematica entry (MEMORY_BACKEND=chroma) plus the GitHub Copilot remote MCP server ghcopilot-warner. To list servers in VS Code: Cmd/Ctrl+Shift+P → MCP: List Servers.

---

Testing with MCP Inspector

npx @modelcontextprotocol/inspector uv run warnerco-mcp
# Opens http://localhost:6274

---

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)

• Website: techtrainertim.com
• GitHub: @timothywarner
• LinkedIn: linkedin.com/in/timothywarner

---

License

MIT License - 2026 Timothy Warner

---

Now go build AI systems that actually remember!