Inspiration
Busy parents often hand children a screen, and the child ends up watching or reading passively. I wanted something closer to a patient helper: a child says what they want to learn, hears the lesson spoken aloud, and answers out loud, with the text on screen to follow along. The parent can then see how it went afterwards.
What it does
TutorTots is a voice-first study helper for young children, built as an MCP server for Alexa+. A child says "teach me about trains." The assistant starts a short lesson written for their age and interests, reads it out section by section, and checks understanding. If the child says "I don't get it," it explains that part again more simply. After the lesson it asks multiple-choice questions and accepts either the letter or the answer in the child's own words. A wrong answer gets a hint and a second try, and then the right answer is shown kindly. A parent can ask how a child is doing, and can tell the assistant about a new child ("this is Amaka, she's 4 and likes animals") so lessons fit them.
How we built it
The core is an MCP server in TypeScript over Streamable HTTP, deployed on Render. It has six tools:
generate_lesson: to generate short lesson on chosen topic.explain_concept: gets called when the child extra explanation or information on a concept or a part of the lesson.get_next_question: used to loop through the question. Everything, lessons and quiz, are returned by the generator, with the quiz hidden. This tool is used to fetched the hidden questions so that it doesn't get served from the start of the lesson.check_answer: used to check if the child's chosen answer is correct. If yes, they move on. If no, they get asked to retry with a hint to help them.get_progress: this is for the parent to check how well the kid is doing in their lesson quizzes. Each child’s age, grade and interests are kept in a household file that the lesson writer reads, so lessons fit the child. I couldn't test on real Alexa+, so I built a web page that imitates the experience, with speech recognition for the child's voice and speech synthesis to read replies aloud. An assistant model stands in for Alexa+ by reading what the child says and choosing which MCP tools to call. The tools use a second model call to write the lessons and explanations. In real Alexa+, the platform would take the place of that first assistant, and the MCP server would stay the same.
Challenges we ran into
- No way to test on real Alexa+: Alexa+ SDK are currently invite only, I switched to the simulated Alexa+ path.
- Models making mistakes with children. The lesson text is requested as structured JSON, and the quiz answers stay on the server, so the assistant never sees them. If the model fails, a ready-made lesson takes over so the session doesn't break or abruptly end, unless it is a server issue.
- Understanding what a child says. Spoken answers are messy, so answer matching works without a model: letters, option text and partial phrases. If I can't tell what was said, the child is asked again instead of being marked wrong.
- Speed. Writing a full lesson with its quiz takes about a minute, which is long for a child. Shorter lessons and a “give me a moment while I write your lesson” message would help.
Accomplishments that we're proud of
A full lesson loop that works by voice: teach, explain again, quiz, hint, score, and progress for the parent. The correct answers never reach the assistant, which keeps the checking honest. The same MCP server works from my web page and from other MCP clients (it currently works as a connector on Claude).
What we learned
I learnt how MCP servers work. I have used them but never built them out so this was a great learning opportunity. In that light, one thing that really stood out to me was the absolute importance of input schema and output schema structuring. I kept running into issues because I either gave it something it was not expecting or I asked for something it doesn't output.
What's next for TutorTots
- We are looking forward to testing on real Alexa+ devices once they're available.
- Basing lessons on curriculum content, such as the Oak National Academy API, instead of relying on the model alone.
- Storing profiles and progress permanently, with sign-in for parents and a simple progress view.
- Looking to cover more subjects and languages.
Log in or sign up for Devpost to join the conversation.