MCP Project Manager Lab

Course: Frontier Engineer Onboarding — Day 1: Claude Code Mastery & Context Engineering Lab: Lab 01 — Build & Deploy a Custom MCP Server (buildable subset) + Context Engineering exercise

What this is

A "Project Manager" MCP server exposing three tools — create_task, list_tasks, update_task — backed by SQLite, plus a well-structured .claude/ configuration directory (settings, custom commands, a hook, and CLAUDE.md).

Related MCP server: Tasks MCP Server

Architecture

Claude Code (MCP Host) <--stdio--> MCP Server (this repo) <--> SQLite
     |
     |-- calls create_task
     |-- calls list_tasks
     |-- calls update_task

• src/db.ts — SQLite persistence layer. Takes a Database.Database instance as a parameter (dependency injection), so it's testable against :memory: without a live server process.

• src/tools.ts — zod schemas + descriptions for the 3 tools.

• src/server.ts — constructs the McpServer and wires each tool to the db layer.

• src/index.ts — stdio entrypoint (creates tasks.db, connects the transport).

Scope & Constraints

This lab is built analytically, matching this workspace's standing no-run/no-deploy convention:

• No deploys. Step 6 of the official lab brief (SSE transport + Railway/Render deployment) is intentionally not built. .mcp.json only contains the local stdio configuration.

• No live runs. Step 5 (opening Claude Code and issuing live prompts against the running server) is not performed. src/index.ts is written but never executed.

• Mocked/local data only. Tests exercise an in-memory (:memory:) SQLite instance as a fixture. No tasks.db file is created or committed.

• TDD without execution. tests/db.test.ts and tests/tools.test.ts are written before their corresponding implementation files (see commit order), but the suite is never run (npm install / npm test are not executed).

Deliverables

• MCP server with 3 tools (create_task, list_tasks, update_task)

• SQLite persistence layer

• .mcp.json local stdio integration

• .claude/ directory: settings.json (permissions + 1 hook), 2 custom commands, well-structured CLAUDE.md

• TDD test suite (written, not executed)

• Deployed to Railway/Render with SSE transport — out of scope (no-deploy constraint)

• Live end-to-end test via Claude Code — out of scope (no-run constraint)

Not built (stretch ideas from the brief, explicitly skipped)

• Extension Challenges: auth on the SSE layer, MCP resource endpoints, a second "Time Tracker" MCP server, prompt templates.

• Extra Exercises A-C: a dedicated database MCP server, an auto-test hook that actually executes test files on edit, a multi-agent git-worktree workflow.

These are skipped, not attempted-and-cut — none require running anything to describe, but building them well would; left out to keep this lab's scope to what's fully buildable and testable without execution.