MCP Project — Deadline / Study schedule tracker

Major assignment for the MCP course (FIT-HCMUS, lecturer Kha Do) — based on the slides MCP — Plug Your Agent Into Everything.

Structure

mcp-homework/
├── server/
│   ├── core.js     ← logic MCP thật: 3 tools + 1 resource + 1 prompt
│   ├── server.js   ← chạy core qua stdio
│   └── http.js      ← chạy core qua Streamable HTTP + API key guard + /health
├── agent/
│   ├── host.js     ← agent = MCP HOST (load config, merge tools, dispatch, use_skill)
│   └── config.json ← khai báo model + các MCP server cần nối + nơi chứa skills
├── skills/
│   └── deadline-review/SKILL.md ← 1 skill orchestrate tool của server
└── README.md

Related MCP server: ScienceTokyoLMS-mcp

Server topic: Deadline / study schedule tracker

• Tools

  • add_deadline({ title, subject, due }) — add a deadline

  • list_deadlines({ subject?, onlyPending? }) — list, with filtering

  • complete_deadline({ id }) — mark as complete, returns isError: true if the id is wrong

• Resource: deadline://list — the full list, read-only

• Prompt: plan_my_week — asks the model to plan the week from existing deadlines

Running each part

1. stdio server (test with Inspector)

npm install
npx @modelcontextprotocol/inspector node server/server.js

2. HTTP server (local, then deploy public)

MCP_KEY=$(openssl rand -hex 32) node server/http.js
# test:
npx @modelcontextprotocol/inspector   # chọn Streamable HTTP, URL http://localhost:3000/mcp,
                                       # header Authorization: Bearer <MCP_KEY>

Deploy (Render/Fly/Railway...): set the MCP_KEY environment variable, expose the PORT, /health is used for the platform's health check.

3. Agent host (using both servers + skill)

cd agent && npm install
# cần Ollama chạy sẵn model qwen3.5:4b (hoặc đổi model/baseURL trong config.json)
node host.js "deadline tuần này của tôi thế nào?"

host.js will: connect the stdio server + HTTP server → merge the tool list → build the skill index from skills/ → run the think/decide/act/observe loop, automatically calling use_skill when the question matches a skill description, then call the corresponding MCP tool.

Self-tested

• Smoke test stdio server: called all 3 tools, read the resource, fetched the prompt, checked isError when completing a non-existent id — all as expected.

• HTTP server: request without key → 401; with correct key → initialize handshake succeeds; /health returns ok.

Still to do before submission

• Deploy the HTTP version to a public platform (Render/Fly/Railway/Cloudflare Workers…), set MCP_KEY, record the URL + key for the grader

• Install Ollama + pull qwen3.5:4b, try running agent/host.js with a real model (the test above only tests the MCP layer, it doesn't call the LLM)

• Record a short demo video (see the "Submission" section below)