Topic 1: Module 6 at a glance, and the setup

By the end of this module, you'll have:

• A host (notes_assistant/host.py) that connects to MCP servers over stdio or Streamable HTTP, converts their tools to the Chat Completions format, and runs a bounded agent loop.
• An agent loop you can test without an API key, with iteration limits, repeated-call detection, and a no-progress extension you wrote yourself.
• Measured numbers for parallel versus sequential tool calls, and timeouts that stop a slow server from freezing the conversation.
• A two-server host (notes plus a lab calendar) where namespacing prevents a real tool-name collision, plus a tool-search meta-tool that cuts tool definitions sent per round by about 86 percent in a 37-tool setup.
• Local rules that decide what runs, what asks, and what never runs, a consent prompt that shows the exact file a write will create, and measured prompt counts showing how a read-only allowlist cuts approval fatigue.
• A clear picture of what a host sees when a server is slow, returns an error, or dies in the middle of a session, and how to recover.

Prerequisites: Modules 1 to 5. You need the notes server from Module 5 (notes_assistant/server.py with build_server(store)), the NoteStore from store.py, and the chat() helper from llm.py. Comfort with async/await and anyio task groups helps.

Where we are: Module 5 built the server side properly: a tested build_server(store) that runs over stdio and HTTP. So far we have only poked it with the Inspector and small client scripts. This module builds the other half: the host that puts a model, a person, and one or more servers in the same room and stays in charge of all three.