MCP Traffic
Main dashboard
Technical proof

Inspiration Most students in India study in a mix of Hindi and English, but voice assistants handle that mix poorly, and study tools live in separate apps. I wanted a coach I can simply talk to: "10 मिनट बाद पढ़ाई याद दिलाओ", "explain relativity", "क्विज़ लो", and have it actually remember my tasks and progress.

What it does Prakash Study Coach exposes 18 tools over MCP so Alexa+ (or any MCP client) can:

set spoken reminders and Pomodoro focus timers that really fire later, manage a to-do list and log studied topics with a progress summary, build a daily study plan for a B.Sc Computer Science and Data Analysis student, explain physics concepts and share space / ISRO facts, run multiple-choice quizzes (the next "A/B/C/D" is treated as the answer), calculate safely (own parser, no eval) and control simulated lights, fan and AC. Every tool replies in Hindi or English depending on the user. An included Agent Skill (SKILL.md) teaches an agent when to call each tool and how to keep replies short enough to be spoken.

How I built it Self-hosted MCP server, Streamable HTTP, spec 2025-11-25, written from the protocol with no dependencies: initialize with version negotiation and Mcp-Session-Id, ping, tools/list, tools/call, session termination via DELETE, Origin validation against DNS rebinding, optional bearer token, tool errors reported as isError results. Simulated Alexa+ web experience: an Echo-style device (listening / thinking / speaking ring), voice in via the Web Speech API and voice out via speech synthesis, plus a live panel showing the real JSON-RPC traffic between the agent and the server. Agent brain with two modes: Claude (via the Anthropic API) chooses tools from the live tools/list; a bilingual offline intent router works without any key. Both call tools through real MCP over HTTP. Tested end to end: a 55-check smoke test starts the server and speaks MCP to it (protocol, validation, security, every tool, the agent in Hindi and English). Challenges Making Hindi intent handling robust (Devanagari normalization, Hinglish phrases), keeping reminders real on a stateless-feeling voice device (server timers plus a polled get_due_reminders tool), and keeping replies short enough for speech.

Accomplishments I'm proud of A complete, protocol-faithful MCP server that runs with a single node server.mjs, a working Agent Skill, and a demo where judges can watch each voice command turn into an actual tools/call.

What I learned How MCP sessions, protocol-version negotiation and Origin checks work in Streamable HTTP, and how to design tool descriptions and a skill so an agent picks the right tool in two languages.

What's next Deploy behind HTTPS and register with Alexa+ using the official developer tooling, add OAuth for per-student accounts, and a real study-analytics view (streaks, weak topics).

Honest scope note Weather is mock data and smart-home devices are simulated. This submission demonstrates the MCP server and the simulated Alexa+ path; it does not claim a deployed Alexa+ skill.