Turn any LLM into an
intelligent tutor.
Tutor MCP is an open-source Model Context Protocol server that turns your AI assistant into an intelligent tutor with persistent progress, session memory, misconceptions, metacognition, and cognitive-science scheduling.
$curl -fsSL https://tutor-mcp.dev/install.sh | sh
Anthropic
OpenAI
Mistral
Geminiget_learner_contextget_next_activityget_memory_stateWelcome back. Last session you were debugging goroutines with a solid mental model for scheduling, but the notes say cancellation still felt blurry.
The runtime is holding two signals together: goroutinesare due for review, and your recent session memory mentions over-confidence on edge cases. We'll do a short prediction step before the exercise, then use that result to decide whether channels should unlock today.
Open Learner Model · Go
Phase · InstructionNext up → a 5-minute goroutines refresher with one cancellation edge case. Ready?
The missing layer is not content.
It is continuity.
Tutor MCP combines hard cognitive signals with human-readable session memory. The next conversation starts from what the learner has mastered, forgotten, misunderstood, felt, and explicitly committed to do next.
Progression
Domains, concepts, prerequisites, mastery, retention, ability, transfer, and review timing stay in SQLite.
Session memory
Each session can leave a factual Markdown trace: affect, concepts touched, salient exchanges, and what to verify next.
Cognitive narrative
Stable learner facts, concept notes, pending observations, misconceptions, and archives give the LLM continuity.
Audit trail
Pedagogical snapshots and decision replay explain why the runtime picked an activity or held a concept back.
Six ways people are
using their tutor.
No course catalog. No textbook. Tell your AI what you want to learn, the tutor handles sequencing, pacing, reviews, and knows when you've actually got it.
Pick up something new from scratch
« I want to learn Spanish before my trip in six weeks. »Tell your AI your goal and your deadline. The tutor maps the path, paces what comes next, and only opens new ground when the basics are solid.
Languages · coding · a new craftBring rusty knowledge back to life
« I studied statistics 8 years ago. »A short diagnostic exposes what's still solid, what's faded, and where the small wrong ideas hide. You revisit only what needs it, not a whole textbook.
Diagnostic · targeted recallPrep for an exam, interview, or certification
« AWS Solutions Architect in three weeks. »Backwards-planned from your test date. Drills focus on your weak spots; reviews peak right before D-day, when memory matters most.
Bar exam · coding interview · AWSFive focused minutes a day
« Just one thing today. »Open the chat in the morning. The tutor knows exactly which concept to revisit, delivers a single calibrated exercise, and keeps the session trace for the next time.
Microlearning · daily habitBuild your own bootcamp or e-learning track
« 12 weeks of data engineering, full-time, starting Monday. »Set the curriculum, the depth, and the pace. The tutor sequences modules on prerequisites, runs daily sessions, and checks mastery between phases. All the structure of a paid bootcamp or e-learning course, none of the cohort calendar.
Self-bootcamp · e-learning · self-pacedTurn material into a course
« Turn this PDF into a six-week course. »Give your LLM the material, then let Tutor MCP keep the concept graph, prerequisites, reviews, and session memory coherent across the course.
Documents · concept graph · memoryLLMs can explain.
Tutor MCP remembers and decides.
The runtime owns the durable learner state and the pedagogical decisions. The LLM stays free to explain, reframe, question, and consolidate narrative memory from the traces it receives.
Tutor MCP
Cognitive signals, phase control, evidence gates, session history, and Markdown learner memory. Predictable. Auditable. No guessing what the learner has actually mastered or what happened last session.
Your LLM
Generates exercises on demand, calibrated by the engine's signals. It writes interpretation briefs, closes sessions, and turns raw state into learner-facing explanations without owning the schedule.
Tells the tutor how confident it should be that you've actually mastered a concept, not just that you happened to answer right today.
Decides when to bring a concept back, using stability and difficulty curves drawn from memory research.
Calibrates each exercise to your current ability, no more "too easy" or "too hard" guesses.
Weighs your wins and losses on each concept to predict how the next attempt is likely to go.
Ensures concepts unlock only when their prerequisites are solid. No castle on sand.
Keeps a deterministic signal for learner ability versus exercise difficulty, separate from the LLM's wording.
Checks whether knowledge moves beyond the training pattern: near, far, debugging, teaching, and creative use.
The engine decides what to teach.
Your AI teaches it ten different ways.
Determinism alone is rigid. LLMs alone drift. Together, they're the first tutor that's both rigorous and infinitely adaptable, built on the model you already trust.
Infinite content
One concept, ten exercises, never the same. No item bank to curate, no editorial team to hire.
A hundred ways to explain
Analogy, theory, story, code, real-world example. The LLM keeps reframing until it actually clicks.
Any language, any tone
Spanish travel phrases or Latin grammar. Formal exam prep or playful drilling. The tutor matches you.
Context-aware by design
When your LLM has a PDF, codebase, note, or image in context, Tutor MCP supplies the learning state that tells it what to do next.
Your tutor.
Your data. Your server.
No telemetry. No SaaS lock-in. No usage caps. Self-host on a 2 GB VPS and own the learner model, memory files, sessions, and backups.
Free forever.
Always yours.
Go 1.25, SQLite, Markdown memory, single binary. Ten minutes from install command to your first session.
Help shape the next
learning primitive.
Tutor MCP is still early. The core is working, but the best version will come from learners, teachers, builders, and MCP users pushing it against real workflows.
If you care about open-source education infrastructure, this is the moment to influence the API, the pedagogy, the docs, and the deployment story.
Test it in the wild
Run Tutor MCP with your real learning goals, then report what feels confusing, brittle, or surprisingly useful.
Improve the tutor brain
Help refine scheduling, mastery checks, misconceptions, transfer challenges, and the learning signals that drive the engine.
Build integrations
Connect more MCP clients, polish provider setup, improve deployment paths, and make self-hosting boringly reliable.
Write better docs
Turn sharp edges into clear guides, examples, diagrams, and recipes that help new users get from zero to first session.
From clone to first
session in ten minutes.
Self-host on your own machine or a small VPS. Connect the AI you already use. Start learning with a runtime that keeps progress, memory, and decisions between sessions.
Run the server
Install the release binary or build the single Go module, set a base64 JWT secret, and expose the MCP endpoint when you want web clients to connect.
$ go build -o tutor-mcp
$ JWT_SECRET=$(openssl rand -base64 32) ./tutor-mcp
Connect your AI
Add Tutor MCP as a custom MCP connector inside the assistant you already pay for. OAuth handles the handshake.
AnthropicOpenAIMistralGeminiTell it what you want to learn
Open your assistant, trigger the tutor, name your goal. From there, the runtime selects activities, records outcomes, closes sessions, and carries the narrative forward.
Questions you might
have.
is a superpower.
LLMs unlocked humanity's knowledge for everyone, but knowing isn't learning. Tutor MCP brings 50 years of intelligent tutoring research back into reach, so anyone, anywhere can build durable, deep understanding from any AI conversation. Self-hosted. Open source. Free, forever.