Converge
Open the app

Week 15 · Season 4 · Will

One plug, many tools

You can build an MCP server that any AI app can plug in: tools agents use well, stateless requests, sign-in with OAuth, and a pause to ask the person when it matters.

Open week 15 in the app

This week

  • The Model Context Protocol: hosts, servers, tools
  • Stateless and remote servers
  • OAuth for tools: tokens, scopes, the door check
  • Elicitation and MCP Apps
  • Tools agents can actually use

How you learn

Before you use an SDK, write the heart of an MCP server by hand: read one JSON-RPC message, answer tools/list and tools/call, and tell a broken message from a tool that tripped.

What you build

FRIDAY gets its own MCP server over stdio with a few sharp tools for your calendar file and one folder of files, plugged into FRIDAY's loop and proved with host-style tests.

Stretch: Put the server online over Streamable HTTP with OAuth: a 401 that points to sign-in, token checks on every request, and no state kept between requests.

What you ship

Project 4: The morning briefing. Every morning FRIDAY gathers your calendar, tasks and notes and briefs you, by itself, within budget.

  • TOOLS.md and a diagram: host, server, tools, data
  • friday_mcp/server.py with host-style tests that pass in CI

Lessons

  1. One plug

    After this deck you can explain the Model Context Protocol (MCP) in plain words: a host app holds the plugs, each server offers a short list of tools, and every request and answer is a small JSON message with a matching id. You can also tell a broken message from a tool error.

    11 minFull access

  2. Stateless and remote

    After this deck you can explain why a remote MCP server keeps no memory between requests (it's stateless), keep state in a database behind a handle the tool returns, and say why a gateway reads the method and tool name from the outside of every request.

    11 minFull access

  3. Who's asking?

    After this deck you can walk through the OAuth sign-in dance a remote MCP server uses (401, find the sign-in server, you approve in a browser, an access token comes back) and write the check a server runs on every request: right issuer, right server, not expired, enough scope.

    11 minFull access

  4. Ask the human, show the screen

    After this deck you can explain elicitation (a server pausing to ask the person one clear question), handle the three answers (accept, decline, cancel), say how a stateless server pauses and retries, and when a small screen (an MCP App) beats words.

    12 minFull access

  5. Tools agents can actually use

    After this deck you can redesign a tool for an agent: fewer, sharper tools with family names, descriptions written like a recipe, short answers with pagination, and errors that say how to fix the call.

    11 minFull access

  6. Go deeper: roads, gifts and guards

    After this deck you can compare the two MCP transports (stdio and Streamable HTTP), say who controls tools, resources and prompts, and name the server-side rules that keep tool servers safe: no token passthrough, check every input, and keep a human in the loop.

    12 minFull accessGo deeper · optional

Labs

Python in the browser, next to the lessons.

Weekend build

Skills you prove

MCP
Build an MCP server that gives any AI app safe, documented tools

Hand-picked official docs, papers and talks.