Quizlar is the learning MCP server. It turns whatever source the user brings — a YouTube lecture, a PDF, a URL, a textbook passage, or a pasted block of notes — into flashcards, then runs
interactive quizzes with FSRS spaced-repetition scheduling so the material actually sticks. Every tool maps one-to-one to something a real learner does: ingest, quiz, track progress,
review what's due.
Built as a voice-first tutor (LiveKit + Deepgram + ElevenLabs); the MCP surface exposes the same primitives the voice agent uses internally, so your agent gets production-grade grading,
STT-aware answer parsing, and the same FSRS scheduler that powers the consumer app.
Three use cases:
1. "Quiz me on X." Call quiz_me(topic) — a composite tool that builds the deck and starts the quiz in one step. Then loop submit_answer → end_quiz. Grading is tier-1 exact match →
phonetic fuzz → short LLM fallback (the same pipeline that ships in the voice product).
2. Study from a YouTube lecture. create_deck_from_youtube pulls the transcript, clusters it into concepts, and generates cards proportional to the video length. Poll get_job_status, then
run the quiz loop.
3. Daily spaced-repetition review. get_study_recommendations returns exactly the cards due under the user's FSRS schedule, prioritised across all decks. Quizlar is the scheduler of record
— your agent executes the plan instead of reinventing one each session.
Auth: Bearer token (sk-qz-<32 chars>) for headless installs, or full OAuth 2.1 / DCR / PKCE for one-click clients (Smithery, Claude Connector). Mint keys at
https://quizlar.app/settings/api-keys.
22 tools total. Voice and text equal-status. Education-first.