OSS-first AI tutor for children that supplements — never replaces — school education.
Local LLM hosting. Curriculum-templated by country and year level. Built-in kid safety from day one.
⚠️ Research preview — supervised pilot only. Mentar is pre-1.0 and not ready for unsupervised use with real children (known safety gaps: no emergency-services signposting, handoff wording not yet professionally reviewed, no PIN gate). Use only with a parent/carer present. See SECURITY.md before running it with a child.
Two real sessions, recorded end-to-end against a locally-run model — same tutor, same safety model, pitched at the year being taught:
| Year 5 Science | Year 12 Maths Methods |
|---|---|
 |
 |
| Multiple-choice question, then an explanation built on a different example (a bird's feathers) so the answer on screen is never given away. | Senior register — "terms with the exact same variable and exponent" — and the worked example ends in the computed answer, not a guess. |
Nothing in those recordings is staged: the questions come from the generators, the marking
from the deterministic verifier, the working from the engine. Only the prose is the model's.
Regenerate them yourself with python3 scripts/record_demo.py (see the script's header).
First run starts small. Only the pilot and practice packs are on out of the box, so a fresh install would open on simple arithmetic — not the Year 12 algebra above. Setup now asks for your child's country and year up front and turns those packs on (157 are shipped, nothing to download) — see Setting it up below. You can change the answer at any time in ⚙️ Settings → Curriculum packs.
A Year 10 two-stage probability question — the child taps Show me how, and after the model's spoken-style explanation, a deterministic worked-example card lays out the branch tree (computed by the engine, not the model — it is always correct):
| Ask for help | Get the working |
|---|---|
 |
 |
The same tutor scales down to a Year 2 counting question, answers wrong answers with a gentle pointer back to the working rather than a red X, and tracks mastery on the curriculum's prerequisite graph:
| Simple end of the range | Wrong answer → gentle retry | Progress map |
|---|---|---|
 |
 |
 |
Two questions, then a lesson. Mentar will not let a child start until a model actually answers, so the first screen is the one that connects it — and the second asks which year your child is in, so the tutor opens on the right material instead of the starter packs.
- Point Mentar at a model. Anything you open before this redirects here. Option A is a
model already running on the machine (Ollama,
llama serve); Option B is one served elsewhere on your network; Option C is a cloud account, behind a written warning and a parent consent step. - Press Save & Connect. It probes the backend there and then — no restart — and tells you what went wrong if it can't reach it, rather than failing later mid-lesson.
- Say which year your child is in. Picking Australia + Year 5 turns on that year's English, Maths and Science. Skippable, re-runnable, and changeable afterwards in Settings.
- Start. The subject cards are the child's home screen from then on.
Recorded against the real app; only the model's prose would differ on your machine. Full instructions, including installing a model from scratch and the tablet setup, are in docs/RUNNING.md.
Mentar is an open-source tutoring framework that lets parents run an AI tutor on their own hardware, with no data leaving the device and no per-seat API fees.
How it teaches — questions first, not lessons. Mentar is not a classroom-style course that lectures and then quizzes. Every session leads with real, checkable questions; teaching happens in response to the child's answers — explanations, worked examples and step-by-step method cards appear when a child is stuck or asks, and an adaptive model (BKT) decides what to ask next from what they have actually demonstrated. Test-to-learn, not lecture-to-test.
The core is three components:
- Template engine — Markdown curriculum files per country + year/grade level, used as learning guidelines. Community-extensible.
- Dialogue framework — Scaffolds tutoring conversations within the bounds of the active curriculum template.
- Safety layer — Content guardrails and age-mode logic baked in, not bolted on. This is the non-negotiable bar the project must clear to justify existing.
Mentar is, candidly, AI-built software. The great majority of the code, tests, and docs in this repo are written by AI agents working under a human maintainer's direction, decisions, and review — not hand-written by a person. In that sense it is close to "vibe coding," even though it follows deliberate engineering discipline: spec-first design, test-driven development (1240+ tests gating changes), design docs before code, versioned prompts, and code review. Those principles raise the quality bar — but they don't change that underlying fact, and we'd rather be upfront about it.
What this means for you:
- The human makes the decisions (scope, safety thresholds, model choices, architecture); the AI executes and advises. Changes are test-gated and reviewed — but the author is AI.
- It has not had a professional, independent audit. The child-safety code and spec are
AI-authored and reviewed by AI plus the maintainer — not by a qualified safeguarding, security,
or child-development professional. As an unfunded OSS project, Mentar cannot commission that
review; it is actively sought pro-bono (there is a ready-to-hand
SAFEGUARDING_REVIEW_PACKET.mdfor any professional willing to contribute) and gates unsupervised mode specifically — seedocs/SAFETY.md§3.5.1. - Treat the project accordingly: carefully built and openly documented, but not yet independently verified. Read the code, run the tests, and do not put it in front of a real child outside a supervised pilot until the open safety items are closed.
Two reasons:
- Privacy — children's data never leaves the device (unless a parent explicitly opts in to a cloud LLM backend — off by default, consent-gated, see
docs/SAFETY.md§4.5). No operator collects it. This is also a major compliance advantage (seecompliance/). - Cost — no per-seat API fees. A parent with a capable laptop or homelab machine pays nothing to run inference.
A paid hosted-inference tier (for non-technical parents) is a planned future bridge, but it carries its own heavier compliance obligations. The OSS local edition stays deliberately data-light by design.
The codebase uses a Python src-layout (src/mentar/); specs and the safety spec live
under docs/ (not in a top-level safety/). See docs/ARCHITECTURE.md
for the authoritative layout.
mentar/
├── curriculum/ # Markdown curriculum templates (concept graphs)
│ ├── _template.md # Authoring format for new curricula
│ ├── itembank/ # Frozen authored items (jsonl) for the pilot graph
│ ├── packs.json # Pack registry (key → template dir + display name)
│ ├── visual_scaffolds/ # Per-topic visual-hint bundle (maths/english/science)
│ └── templates/
│ ├── AU_ACARA/ # Australian Curriculum v9 — maths Y1–12, English Y1–12,
│ │ # science Y1–10, plus senior splits (biology/chemistry/
│ │ # physics/earth-env, literature, maths methods/specialist/
│ │ # general/essential)
│ ├── IN_GENERIC/ # India board-agnostic — maths/English Classes 2–12,
│ │ # science 2–10, senior bio/chem/physics C11–12
│ ├── SG_GENERIC/ # Singapore board-agnostic — maths/English Primary 2 –
│ │ # Secondary 4, science P2–Sec 2, senior bio/chem/physics Sec 3–4
│ ├── US_GENERIC/ # US board-agnostic — maths/English Grades 2–12, science 2–8,
│ │ # then sequenced Biology G9, Chemistry G10, Physics G11
│ ├── _pilot/ # Phase-0 fractions/arithmetic/science pilot graph
│ └── practice/ # Country-agnostic evergreen practice content
├── prompts/ # Versioned prompt templates + prompts/README.md registry (W6.2)
├── config/ # Inference config example + the vetted model roster
├── scripts/ # bootstrap, binary build, item-bank build, backend check, ZIM fetch,
│ │ # release_gate.sh (the full pre-push verification) + git-hooks/
├── src/mentar/ # Python package (src-layout)
│ ├── engine/ # Concept graph (KST), BKT mastery, fringe, probe classifier,
│ │ # item generators (see "How curriculum content is made" below)
│ ├── dialogue/ # Turn-loop controller (session state machine)
│ ├── safety/ # Safety-layer implementation (escalation, output guard, filters)
│ ├── grounding/ # ZIM reader + resolver + data-wrapper (retrieval grounding)
│ ├── inference/ # LLM abstraction layer (swappable backends)
│ ├── eval/ # Deterministic verifiers + model-eval harness
│ ├── db/ # Local SQLite store (schema + access + adapter)
│ ├── tools/ # Template validator, doc-path checker, etc.
│ ├── cli/ # Command-line entry points
│ └── web/ # `mentar serve`'s Flask app (learner + parent views)
├── tests/ # Mirrors the src/ layout
├── docs/ # SPEC, PHASE0(+_STATUS), SAFETY, SESSION_FSM, ARCHITECTURE,
│ # TESTS, CONTENT_LICENSES, PILOT_CONSENT, design/, research/
├── compliance/ # Compliance coverage-status map (points back to docs/)
├── packaging/ # PyInstaller spec + entry point for the single-file binary
├── eval/ # Eval harness configs; datasets/outputs are gitignored.
│ # eval/niah/ holds the needle-in-a-haystack retrieval configs —
│ # run-only: `needlehaystack` is installed on demand, not vendored
├── graphify-out/ # Generated knowledge-graph report (build output, not source)
└── .github/workflows/ # CI + the manual binary build
Every question a child sees today is hand-authored — a parametric formula (maths: e.g. random addition within a range) or a curated fact table (English/science: e.g. a synonym-pair list), written directly in Python and self-validated against hundreds of random draws before shipping. Grounding (the ZIM/retrieval machinery below) is wired to exactly one pack — the original 8-node Phase-0 fractions pilot. All of AU_ACARA / IN_GENERIC / SG_GENERIC / US_GENERIC / practice (every year, every subject) works today with zero ZIM download — there is no correlation between grade level and grounding need in the current build. Grounding, where it exists, only adds a quoted reference passage to an explanation — it never generates a question or decides correctness; that's always the deterministic verifier, never the LLM.
This is a deliberate, proven design for the content it covers (see
docs/EXPLAIN_METHOD_AUDIT.md — a node-by-node audit, last run
in full on 2026-08-11 when the curriculum was 319 nodes; it has not yet been re-run over the
growth to today's 934), but it doesn't
scale to broader/deeper subject coverage by itself — a hybrid direction (keep hand-authoring
where it fits; add a retrieve→extract→verify→freeze pipeline sourced from real ZIM content for
subjects that need it) has been ratified but not yet built — see
docs/design/hybrid_content_architecture.md for the
reasoning, including a live test showing why an ungrounded LLM can't just be asked for facts
directly (a small model got a chemistry equation's final answer right while its shown
reasoning was fabricated nonsense).
Licence note if you ever wire in Khan Academy ZIM content: Khan Academy is CC
BY-NC-SA — fine for this local/personal edition, but the NC (non-commercial) clause
blocks any paid or hosted tier built on it. See
docs/CONTENT_LICENSES.md §3 for the full breakdown before
building anything on top of it.
Codebase map: a Graphify-generated knowledge graph of the code's architecture lives in
docs/CODEBASE_GRAPH.md — contributor material, kept out of this
page so the README stays about the product.
Templates are simple Markdown files that define what topics a child at a given country + year level should be learning. They are guidelines, not scripts — the dialogue framework uses them to keep sessions on-topic and age-appropriate.
Shipping today (157 templates, 934 nodes). Read the counts, not the year ranges — a
year listed below may hold as few as 3–7 topics, and an honest per-year gap report (currently: ZERO AU gaps) against
the real strand structure is generated by python3 -m mentar.tools.audit_curriculum_coverage
(reference: docs/design/curriculum_reference_au.json). Depth build-out is in progress,
senior maths first:
| Pack | Coverage | Alignment |
|---|---|---|
AU_ACARA/ |
Maths Y1–10 (6–10 topics/yr, every reference strand covered) + senior courses: all four courses Y11–12: Essential (14 each), General (12 each), Methods (13–14 incl. the absorbed algebra nodes), Specialist (10 each) — all eight course-years COMPLETE vs the reference · English Y1–10 (5–8 topics/yr, every reference strand covered) + senior courses: Essential English, English (7–8 incl. absorbed nodes), Literature Y11–12 — all COMPLETE vs the reference · Science Y1–10 (4–8 topics/yr, every reference strand covered) · Physics (6–7), Chemistry (6–7), Biology (6 each) + Earth & Environmental Science (5+4) Y11–12 — every senior science subject-year COMPLETE vs the reference | ACARA v9 (CC BY 4.0) for F–10; senior courses: none claimed |
IN_GENERIC/ |
Maths + English Classes 2–12 · Science 2–10 · Physics/Chemistry/Biology 11–12 (now incl. the shared senior-science depth topics) | None claimed — NCERT/CBSE/ICSE licences don't permit it |
SG_GENERIC/ |
Maths + English Primary 2 – Secondary 4 · Science P2 – Sec 2 · Physics/Chemistry/Biology Sec 3–4 | None claimed — Singapore MOE material is all-rights-reserved |
US_GENERIC/ |
Maths + English Grades 2–12 · Science 2–8 · then sequenced: Biology G9, Chemistry G10, Physics G11 | None claimed — Common Core's licence carries a purpose clause + trademark |
_pilot/, practice/ |
Fractions pilot; country-agnostic practice | n/a |
Senior science is split into separate subjects because that is what a senior student actually enrols in — three in parallel (AU/IN/SG) or one per year in sequence (US). Junior years stay one combined "Science", because there they genuinely are combined.
The *_GENERIC packs deliberately claim no syllabus alignment — the level names are display
labels, not assertions about what a country teaches in that year. Licence reasoning per pack:
docs/CONTENT_LICENSES.md §2b.
Anyone can add a new country or year-level template. See curriculum/_template.md for the format.
Kid-safe content blocks and age-appropriate responses are non-negotiable and built in from the start. See docs/SAFETY.md for the full 6-layer spec (implementation lives in src/mentar/safety/).
Key commitments:
- No dark patterns, no compulsive gamification mechanics (legal line under EU AI Act Article 5)
- No emotion recognition or mood inference
- Under-13: parent-mediated mode (parent in the loop, child never alone with AI)
- 13+: more independent with parental oversight available
- Hard block: model must never produce sexual content involving minors
The OSS local edition is data-light by design, which removes most direct developer exposure under COPPA, GDPR-K, and similar frameworks. However, obligations are real and documented.
See compliance/README.md for coverage status — what's mapped, what's incomplete, and where contributors can help.
Mentar is designed to work with smaller OSS models suited to educational dialogue. Low hallucination is critical for a children's tutor. The inference layer is abstracted so users can swap models.
Current evaluation status: see docs/llm-compatibility.md.
Hardware requirements: see docs/hardware-requirements.md.
Full index: docs/index.md. Highlights:
- Spec · Live status · Architecture
- Safety spec (6-layer, non-negotiable) · Pilot consent
- Session state machine · Test plan
- Model evaluation — results, plain-language · roster & plan · eval tooling
- Content licences · Compliance status · Config & grounding sources
- Add or improve a curriculum template under
curriculum/templates/(seecurriculum/_template.mdfor the format) - Improve the safety spec in
docs/SAFETY.md - Fill compliance gaps flagged in
compliance/README.md - Test and document model compatibility in
docs/llm-compatibility.md
Phase 0 pilot underway (single-family, supervised): the end-to-end dialogue loop, local
model evaluation + pick (gemma2:9b), safety pipeline, and learner data model are built and
green, and the stack is now being exercised end-to-end on real family hardware (a base-M1
MacBook). That shakedown is doing its job: it surfaced — and fixed — a setup bug that halved
the generation budget, a reasoning-model failure mode Ollama's think:false cannot switch
off, and timeout/stop-button guarantees that only held on fast hardware. The 4 GB tier has
its first full eval datapoint (qwen3.5:4b — viable, not the recommended minimum;
docs/MODEL.md stays canonical for model choices).
What actually ships today (2026-08-26): 934 concept nodes across 157 curriculum
templates — Australia (ACARA-aligned: maths Y1–12, English Y1–12, science Y1–10, with the
senior years split into biology/chemistry/physics/earth-env, literature, and the four maths
courses) plus three board-agnostic packs that claim no syllabus alignment (IN_GENERIC
Classes 2–12, SG_GENERIC Primary 2–Secondary 4, US_GENERIC Grades 2–12 — each carrying
senior biology/chemistry/physics; see the Curriculum templates table above for the exact
per-subject ranges), a fractions pilot pack, and
country-agnostic practice packs. Every question is generated by hand-authored Python
(parametric formula or curated fact table) and scored by a deterministic verifier — the
model explains, it never decides whether an answer is right.
Known limits, stated plainly: grounding (the ZIM retrieval layer) is wired to the
fractions pilot pack only, not to any of the breadth above; the ratified hybrid content
architecture that closes that gap is designed but not built
(docs/design/hybrid_content_architecture.md).
Unsupervised/independent mode (a child using it with no trusted adult present) stays out of
bounds on the current safety implementation — its safety case needs two things this unfunded OSS
project cannot commission (professionally-validated handoff wording + emergency-services
signposting for the parent-is-the-harm case), and they are sought pro-bono, not funded. The
supervised pilot is not blocked by that: its safety rests on a trusted adult being present.
See docs/SAFETY.md §3.5.1 for the honest position.
Live status tracker: docs/PHASE0_STATUS.md.
AGPL-3.0-only (LICENSE). Free to use — including commercially — provided any
version you distribute or run as a network service keeps its source open under the same terms.
The network clause is deliberate child-safety policy, not ideology: nobody should be able to
offer children a modified Mentar whose safety changes can't be inspected. Commercial licences
outside the AGPL's terms are available from the maintainer; contributions go through a CLA so
that stays possible. Full reasoning, stated honestly: LICENSING.md. Curriculum
content licences are tracked separately in docs/CONTENT_LICENSES.md.
Plot twist: this is secretly an LLM training pipeline. The child is the model, the ZIMs are the corpus, the curriculum is the training schedule, BKT is the eval harness, and bedtime is the compute budget. An LLM built the training rig for a smaller, cuter model. Inception, but with more juice boxes.